The Construction of Optimal Stated Choice Experiments: Theory and Methods (Wiley Series in Probability and Statistics)

The Construction of Optimal Stated Choice Experiments Theory and Methods Deborah J. Street University of Technology, Sy...

Author: Deborah J. Street | Leonie Burgess

45 downloads 680 Views 13MB Size Report

This content was uploaded by our users and we assume good faith they have the permission to share this book. If you own the copyright to this book and it is wrongfully on our website, we offer a simple DMCA procedure to remove your content from our site. Start by pressing the button below!
Report copyright / DMCA form

DOWNLOAD PDF

The Construction of Optimal Stated Choice Experiments Theory and Methods

Deborah J. Street University of Technology, Sydney

Leonie Burgess University ofTechnology, Sydney

111-4

BICLNTSNNIAL

BICKNTENNIAL

WILEY-INTERSCIENCE A John Wiley & Sons, Inc., Publication

This Page Intentionally Left Blank

The Construction of Optimal Stated Choice Experiments

2007

TH E WI LEY 6 ICE NTE NN I AL-KNOWLEDG

E FOR GE N ERAT1ONS

G a c h generation has its unique needs and aspirations. When Charles Wiley first opened his small printing shop in lower Manhattan in 1807, it was a generation of boundless potential searching for an identity. And we were there, helping to define a new American literary tradition. Over half a century later, in the midst of the Second Industrial Revolution, it was a generation focused on building the future. Once again, we were there, supplying the critical scientific, technical, and engineering knowledge that helped frame the world. Throughout the 20th Century, and into the new millennium, nations began to reach out beyond their own borders and a new international community was born. Wiley was there, expanding its operations around the world to enable a global exchange of ideas, opinions, and know-how. For 200 years, Wiley has been an integral part of each generation’s journey, enabling the flow of information and understanding necessary to meet their needs and fulfill their aspirations. Today, bold new technologies are changing the way we live and learn. Wiley will be there, providing you the must-have knowledge you need to imagine new worlds, new possibilities, and new oppottunities.

Generations come and go, but you can always count on Wiley to provide you the knowledge you need, when and where you need it!

c

WILLIAM J. PESCE PRESIDENT AND CHIEF

EXECUTIVE OFFICER

PETER BOOTH WILEY CHAIRMAN OF THE

BOARD

The Construction of Optimal Stated Choice Experiments Theory and Methods

Deborah J. Street University of Technology, Sydney

Leonie Burgess University ofTechnology, Sydney

111-4

BICLNTSNNIAL

BICKNTENNIAL

WILEY-INTERSCIENCE A John Wiley & Sons, Inc., Publication

Copyright 02007 by John Wiley & Sons, Inc. All rights reserved Published by John Wiley & Sons, Inc., Hoboken, New Jersey Published simultaneously in Canada. No part of this publication may be reproduced, stored in a retrieval system, or transmitted in any form or by any means, electronic, mechanical, photocopying, recording, scanning, or otherwise, except as permitted under Section I07 or 108 of the 1976 United States Copyright Act, without either the prior written permission of the Publisher, or authorization through payment o f the appropriate per-copy fee to the Copyright Clearance Center, Inc., 222 Rosewood Drive, Danvers. MA 01923, (978) 750-8400, fax (978) 750-4470, or on the web at www.copyright.com. Requests to the Publisher for permission should be addressed to the Permissions Department, John Wiley & Sons, Inc., 11 1 River Street, Hoboken. NJ 07030, (201) 748-601 I , fax (201) 748-6008, or online at http://www.wiley.comlgo/permission. Limit of LiabilitylDisclaimer of Warranty: While the publisher and author have used their best efforts in preparing this book, they make no representations or warranties with respect to the accuracy or completeness of the contents of this book and specifically disclaim any implied warranties of merchantability or fitness for a particular purpose. No warranty may be created or extended by sales representatives or written sales materials. The advice and strategies contained herein may not be suitable for your situation. You should consult with a professional where appropriate. Neither the publisher nor author shall be liable for any loss o f profit or any other commercial damages, including but not limited to special, incidental, consequential, or other damages. For general information on our other products and services or for technical support, please contact our Customer Care Department within the United States at (800) 762-2974, outside the United States at (317) 572-3993 or fax (317) 572-4002. Wiley also publishes its books in a variety of electronic formats. Some content that appears in print may not be available in electronic format. For information about Wiley products, visit our web site at www.wiley.com. Wiley Bicentennial Logo: Richard J. Pacific0

Library of Congress Cataloging-in-Publication Data: Street, Deborah J. The construction of optimal stated choice experiments : theory and methods l Deborah .IStreet, . Leonie Burgess. p. cm. Includes bibliographical references and index. ISBN 978-0-470-05332-4 1, Combinatorial designs and configurations. 2. Optimal designs (Statistics) I. Burgess, Leonie. 11. Title. QA166.25.S77 2007 51 1'.&dc22 2007011016 Printed in the United States of America. 1 0 9 8 7 6 5 4 3 2 1

CONTENTS

xi

List of Tables

xvii

Preface 1

Typical Stated Choice Experiments 1.1

1.2 1.3 1.4 1 .s

1.6

1.7

1.8

Definitions Binary Response Experiments Forced Choice Experiments The “None” Option A Common Base Option Avoiding Particular Level Combinations 1.6.1 Unrealistic Treatment Cornbinations 1.6.2 Dominating Options Other Issues 1.7.1 Other Designs 1.7.2 Non-mathematical Issues for Stated Preference Choice Experiments 1.7.3 Published Studies Concluding Remarks

1 2 3 5 7 8 9

9 10 11 11 11

12 13

V

vi

CONTENTS

2

Factorial Designs 2.1

2.2

2.3

2.4 2.5 2.6

3

Complete Factorial Designs 2.1.1 2k Designs 2.1.2 3k Designs 2.1.3 Asymmetric Designs 2. I .4 Exercises Regular Fractional Factorial Designs 2.2.1 Two-Level Fractions 2.2.2 Three-Level Fractions 2.2.3 A Brief Introduction to Finite Fields Fractions for Prime-Power Levels 2.2.4 2.2.5 Exercises Irregular Fractions Two Constructions for Symmetric OAs 2.3.1 ; Constructing OA[2k:Z k l , 4 k Z41 2.3.2 2.3.3 Obtaining New Arrays from Old Exercises 2.3.4 Other Useful Designs Tables of Fractional Factorial Designs and Orthogonal Arrays 2.5.1 Exercises References and Comments

The MNL Model and Comparing Designs 3.1

3.2

3.3

3.4

Utility and Choice Probabilities 3.1 . l Utility 3.1.2 Choice Probabilities The Bradley-Terry Model 3.2.1 The Likelihood Function 3.2.2 Maximum Likelihood Estimation 3.2.3 Convergence 3.2.4 Properties of the MLEs Representing Options Using k Attributes 3.2.5 3.2.6 Exercises The MNL Model for Choice Sets of Any Size 3.3.1 Choice Sets of Any Size 3.3.2 Representing Options Using k Attributes 3.3.3 The Assumption of Independence from Irrelevant Alternatives 3.3.4 Exercises Comparing Designs Using Variance Properties to Compare Designs 3.4.1 3.4.2 Structural Properties

15

16 16 19 24 25 27 27 33 37 39 41 41 43 44 46 52 53 55 56 56 57 58 58 59 60 61 62 65 67 70 79 79 79 82 82 83 83 84 88

CONTENTS

3.5

3.4.3 Exercises References and Comments

vii

91 92

Paired Comparison Designs for Binary Attributes

95

4.1

95 97 99

4.2

4.3

Optimal Pairs from the Complete Factorial 4.1.1 The Derivation of the A Matrix Calculation of the Relevant Contrast Matrices 4. I .2 4.1.3 The Model for Main Effects Only The Model for Main Effects and Two-factor Interactions 4.1.4 Exercises 4.1 .5 Small Optimal and Near-optimal Designs for Pairs 4.2.1 The Derivation of the A Matrix 4.2.2 The Model for Main Effects Only 4.2.3 The Model for Main Effects and Two-Factor Interactions 4.2.4 Dominating Options 4.2.5 Exercises References and Comments

100

105 1 I7 118 118

I I9 121 133 134 134

Larger Choice Set Sizes for Binary Attributes

137

5.1

138 138

5.2

5.3

Optirnal Designs from the Complete Factorial 5.1.1 Difference Vectors 5. I .2 The Derivation of the A Mlatrix 5.1.3 The Model for Main Effects Only The Model for Main Effects and Two-Factor Interactions 5.1.4 5.1 .5 Exercises Small Optimal and Near-Optimal Designs for Larger Choice Set Sizes The Model for Main Effects Only 5.2.1 5.2.2 The Model for Main Effects and Two-Factor Interactions 5.2.3 Dominating Options 5.2.4 Exercises References and Comments

143 147 152 159 159

160 163 164 165 165

Designs for Asymmetric Attributes

167

6.1

169 173 174 180 1x0 189 189

6.2 6.3 6.4

Difference Vectors 6. I . 1 Exercises The Derivation of the Information Matrix A 6.2.1 Exercises The Model for Main Effects Only 6.3.1 Exercises Constructing Optimal Designs for Main Effects Only 6.4. I Exercises

197

viii

7

CONTENTS

6.5

The Model for Main Effects and Two-Factor Interactions 6.5.1 Exercises 6.6 References and Comments Appendix 6. A.l Optimal Designs for m = 2 and k = 2 6. A.2 Optimal Designs f o r m = 2 and k = 3 6. A.3 Optimal Designs for m = 2 and k = 4 6. A.4 Optimal Designs f o r m = 2 and k = 5 6. A S Optimal Designs f o r m = 3 and k = 2 6. A.6 Optimal Designs f o r m = 4 and k = 2 6. A.7 Optimal Designs for Symmetric Attributes f o r m = 2

197 209 210 21 1 212 213 217 220 22 I 223 225

Various Topics

227

7.1

7.2

7.3 7.4 7.5 8

Optimal Stated Choice Experiments when All Choice Sets Contain a Specific Option Choice Experiments with a None Option 7.1 .I 7.1.2 Optimal Binary Response Experiments 7.1.3 Common Base Option 7.1.4 Common Base and None Option Optimal Choice Set Size 7.2.1 Main Effects Only for Asymmetric Attributes 7.2.2 Main Effects and Two-Factor interactions for Binary Attributes 7.2.3 Choice Experiments with Choice Sets of Various Sizes 7.2.4 Concluding Comments on Choice Set Size Partial Profiles Choice Experiments Using Prior Point Estimates References and Comments

Practical Techniques for Constructing Choice Experiments 8.1

8.2

8.3

Small Near-Optimal Designs for Main Effects Only 8.1.I Smaller Designs for Examples in Section 6.4 8.1.2 Getting a Starting Design 8.1.3 More on Choosing Generators Small Near-Optimal Designs for Main Effects Plus Two-Factor Interactions 8.2.1 Getting a Starting Design 8.2.2 Designs for Two-Level Attributes 8.2.3 Designs for Attributes with More than Two Levels 8.2.4 Designs for Main Effects plus Some Two-Factor Interactions Other Strategies for Constructing Choice Experiments

228 228 233 234 236 237 237 240 242 243 243 245 246

249 25 1 25 1 256 264 269 269 27 1 272 276 279

CONTENTS

8.4 Comparison of Strategies 8.5 References and Comments Appendix 8. A.l Near-Optimal Choice Sets for C1 = (2 = 2, & = P1 = I , = & = 4, 4, = 8, & = 36, and 7n = 3 for Main Effects Only 8. A.2 Near-Optimal Choice Sets for C1 = P 2 = P3 = L4 = L, = 4, & = 2, P7 = E g = 8, 19 == 24, and in = 3 for Main Effects Only

ix

29 1 293 294 294

296

Bibliography

301

Index

309


LIST OF TABLES

1.1

Attributes and Levels for the Survey to Enhance Breast Screening Participation

I .2

One Option from a Survey about Breast Screening Participation

1.3

Six Attributes to be Used in an Experiment to Compare Pizza Outlets

1.4

One Choice Set in an Experiment to Compare Pizza Outlets

1 .s

1.6

1.7

.x

Attributes and Levels for the Study Examining Preferences for HIV Testing Methods

7

One Choice Set from the Study Examining Preferences for HIV Testing Methods

8

Five Attributes to be Used in an Experiment to Investigate Miscarriage Management Preferences

9

Five Attributes Used to Compare Aspects of Quality of Life

10

2.1

Values of Orthogonal Polynomials for n = 3

22

2.2

A , R, and A R Contrasts for a 32 Factorial

23

2.3

A , R , and A R Contrasts for a s3Factorial

24

2.4

Main Effects Contrasts for a 2 x 3 x 4 Factorial

25

2.5

A Regular 2"'

27

1

Design

xi

Xii

CONTENTS

Designs

28

2.6

Non-overlapping Regular 2’-’

2.7

A 24-1 Design of Resolution 4

29

2.8

Contrasts for the 24 design

30

2.9

Smallest Known 2-Level Designs with Resolution at Least 3

32

2.10


32

2.1 1

A Design with 6 Binary Factors of Resolution 3

33

2.12

A 34-2 Fractional Factorial Design

34

2.13

Pencils for a 3’ Factorial Design

35

2.14

Contrasts for a 3’ Factorial Design

36

2.15

Contrasts for a 34-2 Factorial Design

36

2.16

Smallest Known Regular 3-Level Designs with Resolution at Least 3

37

2.17


38

2.18

The Finite Field with 4 Elements

39

2.19

A Resolution 3 45-3 Fractional Factorial Design

40

2.20

A Resolution 3 Fractional Factorial Design for Seven 3-Level Factors

44

2.21

Generators for 0A[2k;2k1,4ka:41

45

2.22

The OA[64;23,42;4]

46

2.23

A Resolution 3 Fractional Factorial Design for Three 2-Level Factors and Four 3-Level Factors

47

2.24

An OA[16,5,4,2]

48

2.25

An OA[4,3,2,2]

48

2.26

An OA[ 16;2,2,2,4,4,4,4;2]

48

2.27

An OA[ 16,15,2,2]

49

2.28.

An OA[ 12;2,2,2,3;21

50

2.29

An OA[ 16;6,2;2]

51

2.30

An OA[ 12;2,2,6;2]

51

2.31

An OA[ 12;2,2,6;2]without 000

52

2.32

An OA[8,6,2,2]

53

2.33

The Blocks a (7,3,1) BIBD

53

2.34

Some Small Difference Sets

54

2.35

Some Small Difference Families

55

CONTENTS

xiii

3.1

Choice Sets and Choices for Example 3.2.2, with t = 4, s = 5

63

3.2

Estimates of 7ri from All Six Choice Sets

64

3.3

Estimates of

66

3.4

Unconstrained Representation of Treatment Combinations

71

3.5

Constrained Representation of Treatment Combinations

72

3.6

The Fold-over Pairs with X- = 3

79

3.7

Triples of Pairs for k = 2 and the Corresponding D-, A- and EOptimum Values

86

3.8

Two Choice Experiments

88

3.9

A Choice Experiment for the Estimation of Main Effects when There Are 3 Attributes with 2, 3, and 6 Levels which Satisfies the Huber-Zwerina Conditions but for which Main Effects Cannot Be Estimated.

90

An Optimal Choice Experiment for the Estimation of Main Effects when There Are 3 Attributes with 2, 3, and 6 Levels

91

4.1

Six Attributes to Be Used in an Experiment to Compare Pizza Outlets

96

4.2

One Choice Set in an Experiment to Compare Pizza Outlets

96

4.3

The Possible Designs and Corresponding A Matrices for k = 2

99

4.4

All Possible Pairs when k

99

4.5

The 7 Competing Designs for Main Effects Only for Pairs with k = 3

104

4.6

The 7 Competing Designs for Main Effects and Two-Factor Interactions for Pairs with k = 3

108

3.10

7ri

from the First Three Choice Sets Only

=3

4 Binary Attributes

119

4.7

Two Designs with Four Pairs f o r k

4.8

Non-regular OMEPs of Resolution 3

122

4.9

The Pairs from the Complete Factorial when k = 3 and e = ( 0 :1. 1 )

128

4.10

The Pairs from the Fraction i n Example 4.2.4 when e = ( 0 , l . 1 )

128

4.1 1

The Blocks a (7,3,1) BIBD

131

4.12

D-Efficiency and Number of Pairs for Some Constant Difference Choice Pairs

131

5.1

Attributes and Levels for Holiday Packages

138

5.2

One Choice Set for the Possible Holiday Packages

139

5.3

All Possible Triples when k = 3

141

5.4

All Possible Triples when k

143

=

=I

3 Sorted by Difference Vector

xiv

CONTENTS

5.5

All Possible Choice Experiment Designs for Binary Attributes when k = 3andm = 3

152

The Efficiency, for the Estimation of Main Effects plus Two-Factor Interactions, of the 7 Competing Designs when k = 3 and m = 3

158

Optimal Choice Sets for Estimating Main Effects Only for 712 = 5 and k = 9.

163

Near-Optimal Choice Sets for Estimating Main Effects and Two-Factor Interactions f o r m = 3 and k = 4.

165

Attributes and Levels for the Study Examining Preferences for HIV Testing Methods

168


168

6.3

All Possible Choice Sets when m = 3, k = 2,41 = 2, and (2 = 3

170

6.4

All Possible Choice Sets when m = 4, k = 2 , (1 = 2 , and (2 = 3

172

6.5

All Possible Triples when k = 2 , Difference Vector

5.6 5.7 5.8 6.1 6.2

el

= 2, and e 2 = 3 Sorted by

173

All Possible Choice Experiments fork = 2 , PI = 2 , and e 2 = 3 when m=3

190

6.7

Choice Sets when k = 2 , el = 2 , e2 = 3, and m = 3

194

6.8

All Possible Choice Experiment Designs for k = 2, = 2 , and e 2 = 3 when m = 3: Main Effects and Two-Factor Interactions 206

6.9

Values of a"? f o r m = 2 and k = 2

208

6.10

Values of det((7:hI.r) f o r m = 2 and k = 2

208

7.1

Efficiencies of the Four Designs Discussed in Example 7. I . 1

232

7.2

Choice Sets with k = 3 Attributes, el = ez = 2 and P3 = 4

233

7.3

Efficiency of Different Values of m for Five Attributes, Each with Four Levels for Main Effects Only.

239

Efficiency of Different Values of m for Four Asymmetric Attributes for Main Effects Only.

24 1

Efficiency of Different Values of m for 6 Binary Attributes when Estimating Main Effects and Two-factor Interactions.

24 1

Choice Sets of Up to Three Different Sizes for 6 Binary Attributes for Main Effects Only.

243

7.7

Sixteen Attributes Used to Describe Pizzas

244

7.8

(16,20,5,4,1) BIBD

245

6.6

7.4 7.5 7.6

CONTENTS

XV

7.9

The Four Choice Sets from the First Block of the (1 6, 20, 5 , 4 , I)BIBD

245

7.10

The Treatment Combinations and thc Corresponding Assumed

247

8.1

Optimal Choice Sets f o r k = 2, &l == 2, 82 = 3 when m = 3 for Main Effects Only

252

Near-Optimal Choice Sets for k = 2, C1 = &2 = 4 when Main Effects Only

254

8.2

8.3 8.4 8.5 8.6 8.7

Difference Vectors and Sets of Generators for k = 3, and E3 = 6

772

7ri

= 6 for

= 2, (2 = 3,

254

2 x 3 x 61/18 Fractional Factorial Design Obtained from the 3 x 3 x 6//18 by Collapsing One Attribute

255

Near-Optimal Choice Sets for k = 3,41 = 2,02 = 3, and &3 = 6 when m = 3 for Main Effects Only

256

Near-Optimal Choice Sets f o r k = 4, when m, = 3 for Main Effects Only

&I

= 4 = 2, P3 = 3, and

Pj

=6

258

Near-Optimal Choice Sets for k = 4, el = Pz = 2, &3 = 3 , and t-1 = 6 when rn,= 4 for Main Effects Only

259

8.8

Starting Design and Near-Optimal Choice Sets for k = 5 , (1 = T 2 = Y, = 2, and t 4 = &,s = 4 when m = 4 for Main Effects Only 260

8.9

Common Collapsing/Replacement of Attribute Levels

8.10

Choice Sets for k = 5 , t1 = m = 2 for Main Effects Only

8.1 I

e 2 = 173 =

2, and

&4

=

26 1 E5

= 4 when

262

212 x 41/16 Obtained from 45//16 by Expansive Replacement, and Choice Sets for 771 = 3

262

2” x 4//16 Obtained from 21s//16 by Contractive Replacement, and Choice Sets for 777 = 3

263

8.13

A Different 212 x 4//16 and Choice Sets for rn, = 3

264

8.14

Different Designs for k = 17, eq = 2, q = 1 . . . . ,16, and when m = 2 , 3 . 4 , 5 for Main Effects. Only

8.12

017 =

13

266

8.15

45 x 2’ x 81/32 and 48 x 81/82 OMEPs

268

8.16

9 x 45 x 22 x 811288 by Adding Another Attribute to the 45 x 2 2 x 81/32

269

45 x 2 x B3//64 OMEP Obtained from 8g//64 by Collapsing the Levels of 6 Attributes

270

8.18

Fractional Factorial of Resolution 7 for .tq= 2, q = 1 . . . , 7

272

8.19

Near-Optimal Choice Sets for E , = 2, q = 1... . , 7 and rn Main Effects and All Two-Factor Interactions

8.17

~

=

2 for

273

XVi

CONTENTS

Choice Sets for k = 5, C1 = l?, = e3 = e4 = 2 , P5 = 4 when m = 2 for Main Effects and All Two-Factor Interactions

275

Choice Sets f o r k = 3, el = e, = 3, and e3 = 5 when m = 3 for Main Effects and All Two-Factor Interactions

277

Choice Sets for k = 4 Binary Attributes when m = 2 for Main Effects and Interactions AB, AD, and BD

278

Choice Sets for k = 6 Attributes, el = 3, e2 = t3 = e4 = es = 2 , and es = 4 when m = 3 for Main Effects and Interactions AB, AC, AD, AE, and AF

280

8.24

Random Method 1 Choice Sets

28 1

8.25

Random Method 1 Choice Sets: Cnf Matrix

282

8.26

Random Method 2 Choice Sets

283

8.27

Random Method 2 Choice Sets: CAI Matrix

284

8.28

Choice Sets which Satisfy Huber & Zwerina Criteria

285

8.29

Choice Sets which Satisfy Huber & Zwerina Criteria:

8.30

LAfAChoice Sets

288

8.31

Choice Sets from SAS Macros

289

8.32

Choice Sets from SAS Macros: C i t Matrix

289

8.33

Street-Burgess Choice Sets

290

8.34

Street-Burgess Choice Sets: CAIMatrix

290

8.35

Comparison of Construction Methods for Main Effects Only

292

8.36

Comparison of Construction Methods for Main Effects and Two-Factor Interactions 293

8.20 8.21 8.22 8.23

(;.‘A{

Matrix

286

PREFACE

Stated choice experiments are widely used in various areas including marketing, transport, environmental resource economics and public ,welfare analysis. Many aspects of the design of a stated choice experiment are independent of its area of application, however, and the goal of this book is to present constructions for optimal designs for stated choice experiments. Although we will define “optimal” formally later, informally an optimal design is one which gets as much information as; possible from an experiment of a given size . We assume throughout that all the options in each choice set are described by several attributes, and that each attribute has two or more levels. Usually we will assume that all the choice sets i n a particular experiment have the same number of options, although we will relax this constraint in the penultimate chapter. We assume that a multinomial logit model will be used to analyze the results of the stated choice experiment. In the first chapter we describe typical stated choice experiments and give several examples of published choice experiments. We introduce the terminology that we will use throughout the book. In Chapter 2 we define and construct factorial designs. These designs are used in various settings to determine the effect of each of several factors, or attributes, on one or more response variables. Factorial designs are appropriate when discussing the design of stated choice experiments, since in most stated choice experiments the options to be considered are described by attributes each of which can take one of several levels. We show how the effects of each factor can be calculated independently of the other factors in the experiment and we show how the joint effects of two or more factors can be determined. We show the relationship between fractional factorial designs and orthogonal arrays, give

xvii

xviii

PREFACE

relevant constructions and discuss how to use some of the tables of such designs that are available. In Chapter 3 we discuss the use of the multinomial logit (MNL) model to analyze the results of a stated choice experiment. We derive the Bradley-Terry model and extend it to choice sets with more than two options. We show how the attributes that are used to describe the options can be incorporated into the MNL model and hence how to derive the appropriate variance-covariance matrix for the effects of interest. Functions of this matrix are traditionally used to compare designs and we indicate how any two stated choice experiments can be compared so that the “better” design can be determined. We develop the theory for the determination of the optimal stated choice experiment and we show how the optimality value associated with any set of choice sets of any size involving attributes with any number of levels can be calculated. This theory provides a non-subjective way to compare any set of stated choice experiments. We briefly discuss comparing designs using the structural properties of the choice experiments under consideration but as yet there is no firm link between these properties and those of the optimal designs. In the remaining chapters we discuss some specific choice situations in turn. In Chapter 4 we discuss the construction of choice experiments in which all of the attributes describing the options have two levels and in which all choice sets have two options. These are often called paired comparison choice experiments. We find the optimal designs for estimating main effects, and main effects plus two-factor interactions. We get the designs from the complete factorial and show how equally good designs can be constructed from fractional factorials, which were constructed in Chapter 2. In both cases the designs always have known efficiency properties. In Chapter 5 we extend the ideas of the previous chapter to choice sets of any size, although we still retain the restriction that the attributes each have only two levels. We work initially with designs based on thecomplete factorial and then show how to get smaller designs that are just as good from regular fractional factorial designs. In Chapter 6 we extend the ideas of the previous two chapters to construct optimal stated choice designs for any number of attributes with any number of levels using choice sets of any size. We derive the upper bound for D-optimal designs and show how to construct small designs that reach this bound for the estimation of main effects. In this case there are no general constructions for optimal designs for the estimation of main effects and two-factor interactions but we give heuristics that give designs that work well in practice. We give tables of optimal designs for some small situations. In Chapter 7 we briefly consider other important topics in the construction of optimal choice experiments. We look at how to construct optimal designs when there is either a “none of these” option in each choice set or a common base alternative in each choice set. We consider how to design optimal experiments when there are restrictions on the number of attributes that can be different between any two options in a choice set, find the optimal size of the number of options for the choice sets in a choice experiment and look briefly at the use of prior point estimates. The constructions we have given in the previous chapters are not necessarily the easiest way to construct choice experiments so in Chapter 8 we discuss some techniques that we have used in practice to construct optimal or near-optimal choice experiments. We also compare some commonly used strategies for constructing choice experiments. For each chapter we provide references to the mathematical and statistical literature for the constructions, to various literature including the marketing literature and the health services literature for examples of applications of the designs, and we provide a number

PREFACE

xix

of exercises to help the reader test their understanding of the material presented. Some of these exercises provide interesting extensions to the topics discussed in the chapter. Software that allows readers to construct choice sets from a starting design by adding sets of generators is available at http://maths.science.uts.edu.au/maths/wiki/SP~xpts. For these choice sets, or indeed any set ofchoice sets, the software will calculate the information matrix and the corresponding variance-covariance matrix. Our biggest thank you is to Jordan Louviere who introduced the first author to choice experiments a decade ago and who has been a constant source of questions and encouragernent ever since. While writing this manuscript we have benefitted greatly from feedback from various pcople. We would particularly like to thank David Pihlens, Stephen Bush and Amanda Parnis for constructive comments that improved the clarity of the presentation. Each author would like to blame the other author for any mistakes that remain. DEBORAH 1 . STREET A N D

LRONIERIIKGESS


CHAPTER 1

TYPICAL STATED CHOICE EXPERIMENTS

People make choices all the time; some of these decisions are of interest to governments and businesses. Governments might want to model demand for health services in the future or to assess the likely electoral impact of a decision to allow logging in a national park. Businesses want to predict the likely market for new goods and services. Information about choices can be captured from sources like supermarket scanners. However, this “revealed preference” data tells you nothing about products that do not yet exist. Here an experimental approach can help. Such experiments are called “stated preference” or “stated choice” experiments. This book describes the best way to design generic stated preference choice experiments, from a mathematical perspective. Stated choice experiments are widely used in business although often not published. According to the results of a survey sent out to businesses, there were about 1000 commercial applications in the United States in the 1970s and there were about 400 per year in the early 1980s (Cattin and Wittink (1982); Wittink and Cattin (1989)). Wittink et al. (1994) found less extensive use in Europe in the period 1986-1 991, but Hartmann and Sattler (2002) have found that the number of companies using stated choice experiments and the number of experiments conducted had more than doubled in German-speaking Europe by 2001. The range of application areas has also increased and now includes transport (Hensher (1994)), health economics (Bryan and Dolan (2004)), and environmental evaluation (Hanley et al. (2001)),among others. In the rest of this chapter we will define some concepts that we will use throughout the book. We will use published choice experiments from various application areas to illustrate these concepts. These examples will also illustrate the range of issues that needs to be The Consrrucriori or Op~irnnlSmted Choicr 6rperirneril.s. By D.J. Street and L. Burgess Copyright @ 2007 John Wiley 8: Son.;. Inc.

1

2


addressed when designing such an experiment. The mathematical and statistical issues raised will be considered in the remainder of this book.

1.1 DEFINITIONS

Stated choice experiments are easy to describe. A stated choice experiment consists of a set of choice sets. Each choice set consists of two or more options (or alternatives). Each respondent (also called subject) is shown each choice set in turn and asked to choose one of the options presented in the choice set. The number of options in a choice set is called the choice set size. A stated choice or stated preference choice experiment is often called a discrete choice experiment and the abbreviations SP experiment and DCE are very common. We will look at the design of choice experiments for the simplest stated preference situation in this book - the so-called generic stated preference choice experiment. In such an experiment, all options in each choice set are described by the same set of attributes, and each of these attributes can take one level from a set of possible levels. An example involving one choice set might ask members of a group of employees how they will travel to work tomorrow. The five options are {drive, catch a bus, walk, cycle, other} and these five options comprise the choice set. Each respondent will then choose one of these five methods of getting to work. This simple example illustrates the fact that in many choice experiments people are forced to choose one of the options presented. We call such an experiment a forced choice experiment. In this case, being compelled to choose is easy since the respondents were employees (so would be going to work) and every possible way of getting to work was in the choice set since there was an option “other”. Thus the list of options presented was exhaustive. Sometimes a forced choice experiment is used even though the list of options presented is not exhaustive. This is done to try to find out how respondents “trade-off” the different characteristics of the options presented. A simple example is to offer a cheap flight with restrictive check-in times or a more expensive flight where there are fewer restrictions on check-in times. In reality, there might be intermediate choices, but these are not offered in the choice set. However, there are certainly situations where it simply does not make sense to force people to choose. People may well spend several weeks deciding which car to buy and will defer choice on the first few cars that they consider. To allow for this situation, choice experiments include an option variously called “no choice” or “delay choice” or “none of these”. We will just talk about having a none option to cover all of these situations. A related situation arises when there is an option which needs to appear in every choice set. This can happen when respondents are being asked to compare a new treatment with an existing, standard treatment for a medical condition, for instance. We speak then of all choice sets having a common base option. Sometimes just one option is described to respondents who are then asked whether or not they would be prepared to use that good or service. Usually several descriptions are shown to each respondent in turn. This is called a binary response experiment, and in many ways it is the simplest choice experiment of all. It does not allow for the investigation of trade-offs between levels of different attributes, but it gives an indication of combinations of levels that would be acceptable to respondents. When constructing choice sets, it is often best to avoid choice sets where one option is going to be preferred by every respondent. In the flight example above, there will be

BINARY RESPONSE EXPERIMENTS

3

respondents who prefer to save money and so put up with the restrictive check-in conditions, and there will be respondents for whom a more relaxed attitude to the check-in time will be very important. I t seems obvious, though, that a cheap flight with the relaxed check-in conditions would be preferred by all respondents. An option which is preferred by all respondents is called a dominnting or dominant option, and it is often important to be able to design choice experiments where it is less likely that there are choice sets in which any option dominates (or where there is an option that is dominated by all others in the choice set). There is a discussion about dominating options with some discussion of earlier work in Huber and Zwerina (1996). When constructing options which are described by two or more attributes, i t can be necessary to avoid unrealistic combinations of attribute levels. For example, when describing a health state and asking respondents whether they think they would want a hip replacement if they were i n this health state, it would be unrealistic to describe a state in which the person had constant pain but could easily walk 5 kilometers. Throughout this book, we will only consider situations where the options in the experiment can be described by several different attributes. Each attribute has two or more levels. For the flight example we have been describing, the options have two attributes, the cost and the check-in conditions. In general. attributes need to have levels that are plausible and that are varied over a relevant range. For example, health insurance plans can be described by maximum cost to the subscriber per hospital stay, whether or not visits to the dentist are covered, whether or not visits to the physiotherapist are covered, and so on. Although attributes like cost are continuous, in the choice experiment setting we choose a few different costs and use these as discrete levels for the attribute. Thus we do not consider continuous attributes in this book. Finally, we stress that we will be talking about generic- stated preference choice experiments throughout this book. We do not consider the construction of optimal designs when the options are labeled, perhaps by brand or perhaps by type of transport, say, and hence the attributes, and the levels, depend on the label. We only consider designs that are analyzed using the M N L model.

1.2

BINARY RESPONSE EXPERIMENTS

As we said above, in binary response experiments the respondents are shown a description of a good or service, and they are asked whether they would be interested in buying or using that good or service. For each option they are shown, they answer “yes” or “no”. One published example o f a binary response experiment appears i n Gerard et al. (2003). This study was carried out to develop strategies that were likely to increase the participation rates in breast screening programs. The goal of a breast screening program is to achieve a target participation rate across the relevant population since then there should be a reduction in breast cancer mortality across that population. To get this participation rate requires that women participate at the recommended screening rate. The aim of the study described in Gerard et al. (2003) was to “identify attributes of service delivery that eligible screenees value most and over which decision makers have control”. The attributes and levels used in the study are given in Table I . I . Given these attributes and levels, what does a respondent actually see? The respondent sees a number of options, like the one in Table I .2, and just has to answer the question. In this particular survey each respondent saw 16 options (invitations) and so answered the question about each of these I6 different possible invitations i n turn.

4


Table 1.1

Attributes and Levels for the Survey to Enhance Breast Screening Participation

Attributes

Attribute Levels

Method of inviting women for screening

Personal reminder letter Personal reminder letter and recommendation by your GP Media campaign Recommendation from family/friends

Information included with invitation

No information sheet Sheet about the procedure, benefits and risks of breast screening

Time to wait for an appointment

1 week 4 weeks

Choice of appointment times

Usual office hours Usual office hours, one evening per week Saturday morning

Time spent traveling

Not more than 20 minutes Between 20 and 40 minutes Between 40 and 60 minutes Between 1 and 2 hours

How staff at the service relate to you

Welcoming manner Reserved manner

Attention paid to privacy

Private changing area Open changing area

~

Time spent waiting for mammogram

20 minutes 30 minutes 40 minutes 50 minutes

Time to notification of results

8 working days 10 working days 12 working days 14 working days

Level of accuracy of the screening test

70% 80% 90% 100%

FORCED CHOICE EXPERIMENTS

5

Table 1.2 One Option from a Survey about Breast Screening Participation Screening Service ~

How are you informed

Information provided with invitation Wait for an appointment Appointment choices Time spent traveling (one way) How staff relate to you Changing area Time spent having screen Time waiting for results Accuracy of the results

Personal reminder letter No information sheet 4 weeks Usual office hours and one evening per week Not more than 20 minutes Reserved manner Private changing area 40 minutes 10 working days 90%

Imagine that your next invitation to be screened is approaching. Would you choose to attend the screening service described above? (rick one only) Yes 0

No 0

The statistical question here is: “which options (that is, combinations of attribute levels) should we be showing to respondents so that we can decide which of these attributes, i f any, is important, and whether there are any pairs of attribute levels which jointly influence the decision to participate in the breast screening service?” The design of such informative binary response designs is described in Section 7.1.2.

1.3 FORCED CHOICE EXPERIMENTS

In a forced choice experiment, each respondent is shown a number of choice sets i n turn and asked to choose the best option from each choice set. There is no opportunity to avoid making a choice in each choice set. Severin (2000)investigated which attributes made take-out pizza outlets more attractive. In her first experiment, she used the six attributes in Table 1.3 with the levels indicated. A sample choice set for an experiment looking at these six attributes describing take-out pizza outlets is given in ‘Table 1.4. There are three things to observe here. The first is that all the attributes have two levels; an attribute with only two levels is called a bii?flqattribute, and it is often easier to design small, but informative, experiments when all the attributes are binary. We focus on designs for binary attributes in Chapters 4 and 5 . The second is that the question has been phrased so that the respondents are asked to imagine that the two choices presented to them are the last two options that they are considering in their search for a take-out pizza outlet. This assumption means that the respondents are naturally in a setting where it does not make sense not to choose an option, and so they are forced to make a selection even though the options presented are not exhaustive.

6


Finally, observe that each respondent has been shown only two options and has been asked to state which one is preferred. While it is very common to present only two options in each choice set, it is not necessarily the best choice set size to use; see Section 7.2 for a discussion about the statistically optimal choice set size. Larger choice sets do place more cognitive demands on respondents, and this is discussed in Iyengar and Lepper (2000), Schwartz et al. (2002), and Iyengar et al. (2004). Table 1.3 Six Attributes to be Used in an Experiment to Compare Pizza Outlets Attributes

Attribute Levels

Pizza type

Traditional Gourmet

Type of Crust

Thick Thin

Ingredients

All fresh Some tinned

Size

Small only Three sizes

Price

$17 $13

Delivery time

30 minutes 45 minutes

Table 1.4 One Choice Set in an Experiment to Compare Pizza Outlets

Pizza type Type of crust Ingredients Size Price Delivery time

Outlet A

Outlet B

Traditional Thick All fresh Small only $17 30 minutes

Gourmet Thin Some tinned Small only $13 30 minutes

Suppose that you have already narrowed down your choice of take-out pizza outlet to the two alternatives above. Which of these two would you choose? (tick one onlv) Outlet A 0 Outlet B 0

Most forced choice experiments do not use only binary attributes. Chapter 6 deals with the construction of forced choice experiments for attributes with any number of levels. For example, Maddalaet al. (2002) used 6 attributes with 3,4,5,3,s. and 2 levels, respectively,

THE”N0NE OPTION

7

in a choice experiment examining preferences for HIV testing methods. The attributes, together with the attribute levels, are given in Table 1.5, and one choice set from the study is given in Table 1.6. Each respondent was presented with 1 1 choice sets and for each of these was asked to choose one of two options. As the respondents were all surveyed at HIV testing locations a forced choice experiment was appropriate. Tahle 1.5

Attributes and Levels for the Study Examining Preferences for HIV Testing

Methods

Attribute

Attribute Levels

Location

Doctor’s office Public clinic Home

Price

$0 $10 $50 $100

Sample collection

Draw blood Swab mouth Urine sample Prick finger

Timeliness/iiccuracy

Results in 1-2 weeks; almost always accurate Immediate results; almost always accurate Immediate results; less accurate

Privacy/anonymity

Only you know; not linked Phones; not linked In person: not linked Phone; linked In person; linked

Counseling

Talk to a counselor Read brochure then talk to counselor

Both of the experiments discussed above used only six attributes. Hartmann and Sattler (2002) found that about 75% of commercially conducted stated choice experiments used 6 or fewer attributes and they speculated that this might be because commonly available software often used to generate choice experiments would not allow more than 6 attributes. However, there are choice experiments where many attributes are used; see Section 1.7.2. 1.4 THE “NONE” OPTION As we said above, sometimes it does not make sense to compel people to choose one of the options in a choice set and so some choice experiments include an option variously called “no choice” or “delay choice” or “none of these” in each choice set. Often an existing forced choice experiment can be easily modified to include an option not to choose. For instance, in the pizza outlet experiment described in the previous section,

8


Table 1.6


Attribute

Option A

Option B

Location Price Sample collection Timeliness/accuracy

Doctor’s office

Public clinic

Privacy/anonymity Counselling

$100

$10

Swab mouth Results in 1-2 weeks; almost always accurate In person; not linked Talk to a counselor

Urine sample Immediate results; less accurate Only you know; not linked Read brochure then talk to counselor

Which of these two testing methods would you choose? (tick one on!.^) Option A 0 Option B 0

we could change the question to ask: “Suppose that you have decided to get a take-out meal. Which of these pizza outlets would you select, or would you go somewhere else?’ Dhar (1997) looks at the situations in which consumers find it hard to choose and so will opt to defer choice if they can. Haaijer et al. (2001) summarize his results by saying “respondents may choose the no-choice when none of the alternatives appears to be attractive, or when the decision-maker expects to find better alternatives by continuing to search. ...adding an attractive alternative to an already attractive choice set increases the preference of the no-choice option and adding an unattractive alternative to the choice set decreases the preference of the no-choice.” In Section 7.1 .l,we discuss how to design good designs when there is a “none of these” option in each choice set.

1.5 A COMMON BASE OPTION Some choice experiments have a common (or base) option, sometimes called constant comparator, in each choice set, together with one or more other options. This is often done so that the current situation can be compared to other possibilities. A typical example arises in medicine when the standard treatment option can be compared to a number of possible alternative treatment options. Ryan and Hughes (1997) questioned women about various possible alternatives to the surgical removal of the product of conception after a miscarriage (note that some such treatment is essential after a miscarriage). They identified the attributes and levels given in Table 1.7 as being appropriate. In the choice experiment, the common base, which is the current treatment option of surgical management, was described as “having a low level of pain, requiring 1 day and 0 nights in hospital, taking 3-4 days to return to normal activities, costing $350, and there would be complications post-surgery”. This raises the question of whether it is sensible to have an option in which it is known with certainty that there will be complications; it might have made more sense to talk of the probability of complications given a particular treatment. This interesting non-mathematical issue will not be addressed in this book.

AVOIDING PARTICULAR LEVEL COMBINATIONS

9

Table 1.7 Five Attributes to he Used in an Experiment to Investigate Miscarriage Management Preferences

Attributes

Attribute Levels

Level of pain

Low Moderate Severe

Time in hospital

I day and 0 nights 2 days and I night 3 days and 2 nights 4 days and 3 nights

Time taken to return to normal activities

1-2 days 3 4 days 5-6 days More than 7 days

Cost to you of‘treatment

$100

$200 $350 $500 Complications following treatment

Yes NO

Having a common option in all choice sets is not as good as allowing all the options to be different from one choice set to another but when a common base is appropriate we show how to design as well as possible for this setting in Section 7.1.3.

1.6 AVOIDING PARTICULAR LEVEL COMBINATIONS Sometimes a set of level combinations of at least two of the attributes is unrealistic and sometimes a set of level combinations is clearly the best for all respondents and so will always he chosen. We discuss examples of each of these situations here. We give some ideas for how to design choice experiments when these circumstances pertain i n Chapters 4, 5, and 8.

1.6.1

Unrealistic Treatment Combinations

To illustrate this idea consider the descriptions of 5 attributes describing health states devised by EuroQol; see EuroQoL (2006). These attributes and levels are given in Table 1.8 and are used to describe health states for various purposes. In the context of a stated preference choice experiment we might describe two health states and ask respondents which one they prefer. But even a quick look at the levels shows that some combinations of attribute levels do not make sense. A health state in which a person is “Confined to bed” in the mobility attribute is not going to be able to be linked with “No problems with self-care” in the self-care attrihute. Thus it is necessary to determine the level Combinations that are unwolistir before using

10


choice sets that are generated for these attributes using the techniques developed in later chapters.

Table 1.8 Five Attributes Used to Compare Aspects of Quality of Life Attributes

Attribute Levels

Mobility

No problems in walking about Some problems in walking about Confined to bed

Self-care

No problems with self-care Some problems with self-care Unable to wash or dress one’s self

~~~~

Usual Activities (e.g., work, study housework)

No problems with performing one’s usual activities Some problems with performing one’s usual activities Unable to perform one’s usual activities

Pain/Discomfort

No pain or discomfort Moderate pain or discomfort Extreme pain or discomfort

AnxietyDepression

Not anxious or depressed Moderately anxious or depressed Extremely anxious or depressed

1.6.2 Dominating Options Many attributes can have ordered levels in the sense that all respondents agree on the same ordering of the levels for the attribute. In the levels presented in Table 1.8 it is clear that in every attribute the levels go from the best to the worst. Thus a choice set that asks people to choose between the health state (No problems in walking about, No problems with self-care, No problems with performing one’s usual activities, No pain or discomfort, Not anxious or depressed) and any other health state is not going to give any information all respondents will choose the first health state. We say that the first health state dominates the other possible health states. It is possible to have a choice set in which one option dominates the other options in the choice set even though the option is not one that dominates all others in the complete set of level combinations. So if we use 0, 1, and 2 to represent the levels for each of the attributes in Table I .8 then the best health state overall is 00000. But a choice set that contains (001 11,01222, 1221I), for example, has a dominating option since 001 11 has at least as good a level on every attribute as the other two options in the choice set. It is not clear, however, whether 01 222 or 1221 1 would be preferred since it is not necessarily true that the same utility values apply to the same levels of different attributes. We discuss ways of avoiding choice sets with dominating options in Chapters 4,5 , and 8.

OTHER ISSUES

11

1.7 OTHER ISSUES

In this chapter we have discussed a number of different types of choice experiments that have been published in the literature. We have indicated that, in later chapters, we will describe how to construct optimal designs for binary responses, for forced choice stated preference experiments, for stated choice experiments where a “none” option is included in every choice set, and for stated choice experiments where a common base option is included in every choice set. In all cases we consider only generic options analyzed using the MNL model. In this section we want to mention a couple of designs that we will not be constructing and to discuss briefly some non-mathematical issues that need to be considered when designing choice experiments. 1.7.1 Other Designs

In some experiments, options are described not only by attribute levels but also by a brand name or label; for example, the name of the airline that is providing the flight. Such experiments are said to have branded or labeled options (or alternatives), and these have alternative .specific attributes. In many situations, people choose from the options that are available at the time they make their choice rather than deferring choosing until some other option is available. Experimentally, we can model this by having a two-stage design process. We have a design which says what options are available and another design that determines the specific options to present given what is to be available. Such designs are termed maiIatii/ity designs. We will not be discussing the construction of designs for branded alternatives or for availability experiments in this book. The interested reader is referred to Louviere et al. (2000). 1.7.2

Non-mathematical Issues for Stated Preference Choice Experiments

The first issue, and one that we have alluded to in Sections 1.3 and 1.4, is the question of task complexity and thus of respondent efficiency. If a task is too complicated (perhaps because there are too many attributes being used to describe the options in a choice set or because there are too many options in each choice set), then the results from a choice experiment are likely to he more variable than expected. Aspects of this problem have been investigated by various authors, and several relevant references are discussed i n lyengar and Lepper (2000). Louviere et al. (2007) found that completion rates are high even for what would be considered large choice experiments in terms of the number of attributes and the number of options. A related issue concerns the number of choice sets that respondents can reasonably be expected to complete. If there are too many choice sets, then respondents may well become tired and give more variable results over the course of the experiment. This situation has been investigated by Brazell and Louviere (1995) and Holling et al. (1998). The discrete choice experiment task needs to be thought about in context. Are the choices one time only or are they repeated? How important is the outcome of the decision? (In a medical setting you could be asking people to think about life-or-death decisions.) How familiar is the context? The choice of a health insurance provider may be familiar while other choices, such as for liver transplant services, may require the provision of detailed information in the experiment so that respondents can make an informed choice.

12


The attributes that are used to describe the options in a choice experiment need to be appropriate and plausible, and the combinations that are presented in the experiment need to be realistic; otherwise, respondents may take the task less seriously or be confused by it. The mechanics of setting up a stated preference experiment are outlined in Chapter 9 of Louviere et al. (2000). Dillman and Bowker (2001) discuss many aspects of both mail and internet surveys. There are various sources of bias that have been identified as sometimes occurring in choice experiments. One is what has been termed qfirmafion bias when respondents choose responses to be consistent with what they feel the study objectives are. This is why experimenters sometimes include attributes that are not of immediate interest to mask the main attributes that are under investigation. A second possible source of bias is called rationalization bias, where responses are given that justify the actual behavior. This serves to reduce cognitive dissonance for the respondents. A third possible source of bias results from the fact that there are no transaction costs associated with choices in a stated preference study. Some respondents try to respond in a way that they believe will influence the chance of, or the magnitude of, changes in the real world. This is termed strategic or policy response bias. Finally, people may not be prepared to indicate preferences which they feel are socially unacceptable or politically incorrect. These terms are defined and used in Walker et al. (2002). Other sources of bias can be related to the actual topic under investigation. Carlsson (2003) was investigating business passengers preferences, and some of his options were more environmentally friendly, but more expensive, than other options, and he spoke of respondents perhaps aiming to have a “warm glow” or “purchasing moral satisfaction” when making choices. Severin (2000) has shown, in a paired comparison experiment, that if there are a large number of attributes describing each option, then respondents find the task more difficult and she has suggested that about 8 or 9 attributes seem to be effectively processed; see her thesis for more details. For a discussion of the role of cognitive complexity in the design of choice experiments, readers are referred to DeShazo and Fermo (2002) and Arentze et al. (2003) and references cited therein. These and other psychological and cognitive issues are beyond the scope of this book, and we refer the reader interested in such issues elsewhere. The papers by Iyengar and Lepper (2000) and Schwartz et al. (2002) and the references cited therein provide a good starting point to find out more about these issues.

1.7.3

Published Studies

As we have said before, stated preference choice experiments are used in many areas. Here we give references to a few published studies, together with a very brief indication of the question being investigated and the type of design being used. Chakraborty et al. (1994) describe a choice experiment to investigate how consumers choose health insurance. As well as the actual company offering the insurance, 23 other attributes were used to describe health insurance plans. Respondents were presented with choice sets with 4 options i n each and asked to indicate their preferred plan. Hanley et al. (2001 ) describe a stated preference study to investigate demand for climbing in Scotland. Each choice set contained two possible climbs and a “neither of these” option. They also give a table with details of about 10 other studies that used DCEs to investigate questions in environmental evaluation.

CONCLUDING REMARKS

13

Kemperman et al. (2000) used an availability design to decide which of four types of theme parks would be available to respondents in each of spring and summer. Each choice set contained four theme parks and a “none of these” option. McKenzie et al. (2001 ) describe a study in which five attributes that are commonly used to describe the symptoms of asthma were included and three levels for each of these chosen. Patients with moderate to severe asthma were shown “pairs of scenarios characterized by different combinations of asthma symptoms”, and were asked which of the two scenarios they thought would be better to have or whether they felt there was no difference. Ryan and Gerard (2003) and Bryan and Dolan (2004) give examples of the use of DCEs in the health economics context. Scarpa et al. (2004) discuss experiments to “characterise the preference for fifteen different attributes related to water provision”. An example of a choice experiment that involved labeled options is given in Tayyaran et al. (2003). The authors were interested in investigating whether telecommuting and intelligent transportation systems had an impact on residential location. Each choice set contained three residential options as well as a “none of these” option. The residential options were “branded” as central cities, ,first-tiersatellite nodes, and second-tier satellite nodes. Each residential option was described by 7 attributes, 4 of which were the same for all of the locations. Walker et al. (2002) used two stated choice experiments to model tenants’ choices in the public rental market. 1.8 CONCLUDING REMARKS

In this chapter we have seen that there are a number of areas where stated choice experiments have been applied and that there are a number of issues, both mathematical and non-mathematical, which need to be considered in the construction of the best possible experiments for a given situation. In the remainder of this book we look at the mathematical issues that need to be considered to design a good generic stated choice experiment to be analyzed using the M N L model. In the next chapter, we collect a number of results about factorial designs which are intimately connected with the representation of options by attributes. We follow this with a discussion about parameter estimators and their distribution. Over the following three chapters we show how to get the best designs for any number of attributes, each attribute being allowed to have any number of levels and with choice sets of any size. In the penultimate chapter we consider how to construct good designs for other situations, such as the inclusion of a “none” option in every choice set. The final chapter illustrates the application of the results in the book to the construction o f a number of experiments that we have designed in the last five years.


CHAPTER 2

FACTORIAL DESIGNS

Comparative experiments are carried out to compare the effects of two or more treatments on some response variable (or variables). They were developed in an agricultural setting, and often the effects of several factors on the yield of a crop, say, were investigated simultaneously. Such experiments are calledfactnrial experinients and were introduced by Sir Ronald Fisher in the 1920s. Suppose that an experiment is designed to investigate the effect of I; factors on the yield of a crop. Then the treatments of interest are in fact combinations of levels of each of the k factors under investigation. We refer to these as treatment combinations. If the qth factor k tqpossible treatment combinations, but often only has t qlevels, then there are L = a subset of these are actually used in the experiment. In this chapter we will show the link between the treatment combinations in a factorial design and the options used in a stated preference choice experiment. We will review the constructions that exist for obtaining subsets of the treatment combinations from a factorial design so that specific effects of interest can still be estimated. While these constructions were developed to provide good designs when the results of the experiment are analyzed using a linear model with normal errors, it turns out that these constructions are still useful when the multinomial logit model is used. Thus the results developed here form the basis for much of the rest of the book. Throughout this chapter, we will assume that yi,ri2,,..,2b. is the response with the first factor at level i , the second factor at level j , and so on, until the final factor is at level q in a traditional comparative experiment. (In choice experiments the word attribute is used instead of the word factor. In this chapter we use the word “factor,” since we are

n,=,

Thr Corislnrc/ioti ofOp/imn/Sinred Choice Ekperimrnm By D. J . Street and L. Burgess Copyright @ 2007 John Wiley & Son\. Inc.

15

16

FACTORIAL DESIGNS

summarizing constructions from the statistical literature. In later chapters we will use the word “attribute” instead, to make the results more consistent with the use in various application areas.)

2.1 COMPLETE FACTORIAL DESIGNS A complete factorial design is one in which each of the possible level combinations appears at least once. If all of the factors have the same number of levels, then the factorial design is said to be symmetric; otherwise it is asymmetric. Suppose that there are k factors and that each of the factors has 4 levels. Then the factorial design is said to be a symmetric design with 4 levels, and we talk about an ek factorial design. The smallest symmetric factorial design is a design in which two factors each have two levels, a 22 design. We will represent the levels of an 4 level factor by 0. 1 , . . . ,4 - 1 . EXAMPLE 2.1.1. The combinations of factor levels in a 23 factorial design are 000,001,010,011, 100, 101, ll0,and 111.

In an asymmetric factorial design, the factors may have different numbers of levels. If factor q has P, levels, then we speak of an PI x 42 x . . . x e k factorial design. EXAMPLE 2.1.2.

The combinations of factor levels in a 3 x 4 factorial design are 00,01,02, 03, 10, 1 1 , 12, 13, 20, 2 I , 22, and 23. 0

2.1.1 2k Designs Designs in which all factors have only two levels are the most commonly used factorial designs. Typically, the two levels are chosen to be the largest and smallest levels that are deemed to be plausible for that attribute (factor). For example, in an experiment to investigate treatment preferences for asthma, one of the attributes might be sleep disturbance with levels no disturbance and woken more than 5 times per night. In an experiment describing employment conditions, one of the attributes might be amount ofannual leave with levels 2 weeks per year and 6 weeks paid leave and up to 4 weeks unpaid leave per year. In an experiment to compare plane flights, one of the attributes might be iny7ight service with levels beverages and hot meals. 2.7.1.7 Main Effects As a result of conducting the experiment, we would like to be able to estimate the effect of each of the factors, individually, on the response. This is termed the main effect of that factor. If we fix the levels of all factors except one, and then look at the difference in response between the high level and the low level of this one factor, this difference in response is called the simple effect of the factor at the particular levels of the other factors. The main effect of a factor is the average of all the simple effects for that factor. mEXAMPLE2.1.3.

Suppose that there are k = 3 factors, A , B, and C, each with two levels. Suppose that a complete experiment has been carried out. Then there are four simple effects to calculate

COMPLETE FACTORIAL DESIGNS

17

for each of the three factors. Consider factor A . Then the simple effect of A at level 0 for B and level 0 for C' is given by Y l O O - YOOO.

The simple effect of A at level 0 for B and level 1 for C is given by Y l O l - YOOl.

The simple effect of A at level 1 for B and level 0 for C is given by Y l l O - YOlO.

Finally. the simple effect of A at level 1 for B and level 1 for C is given by Yl11 - Y O 1 1

The main effect of A is the average of these four simple effects and is given by 1 ;(VIOO

-

ynoo

+ ~ i o -i Y O O ~+ ~ i i -o Y O ~ O+ ~ i i -i ~ 0 1 1 ) .

0

It is often easier to describe the main effect of a factor as the difference between the average response to that factor when it is at level 1 and the average response when it is at level 0. To write this mathematically, we let the treatment combinations be represented by binary k-tuples ( T ~ : c z . zk), . where zg = 0 or I , q = 1.. . . k . Then we could write the main effect of the qth factor as ~

2.1.1.2 interaction Effects Two factors are said to interact if the effect of one of the factors on the response depends on the level of the other factor. Formally the twofactor interaction efect of factors A and B is defined as the average difference between the simple effect of A at level 1 of B and the simple effect of il at level 0 of B. EXAMPLE 2.1.4.

Suppose that k = 2. Then the simple effect of A at level 1 of B is ( ~ 1 1- Y o l ) and the simple effect of A at level 0 of B is (y10 - ~ 0 0 ) .The interaction effect of A and R , denoted by A B , is the average difference between these two simple effects, that is,

In general, a two-factor interaction is

18

FACTORIAL DESIGNS

Higher-order interactions are defined recursively. Thus the three-factor interaction between factors A, B, and C i s defined as the average difference between the A B interaction at level 1 of C and the A B interaction at level 0 of C . mEXAMPLE 2.1.5. Suppose that k = 4. Then the A B interaction at level 1 of C is 1

4 (Y1111 + Yo011

- (Yo111

+ Y l O l l ) + Y l l l O + Yo010 - ( Y O l S O + Y l O l O ) )

The AB interaction at level 0 of C is 1

+ Yo001

i(Yll01

- (Yo101

+ Y l O O l ) + Y l l O O + Yoooo - ( Y O 1 0 0 + Y l O O O ) ) .

The average difference between these is

ABC=

;{ [i

(Y1111

-

1 --

+ Yo011 - ( Y O 1 1 1 + Y l O l l ) + Y l l l O + yo010 - (Yo110 + Y l O l O ) )

"

4 ( Y l l O l + Yo001 - ( Y O 1 0 1 + Y l O O l ) + Y l l O O + yoooo - ( Y O 1 0 0 + Y l O O O ) )

1

+ Yo011 + Y l l l O + Yo010 + Yo101 + Y l O O l + Yo100 + Y l O O O ) (Yo111 + Y l O l l + Y O 1 1 0 + Y l O S O + Y l l O l + Yo001 + Y l l O O + Yoooo)) .

((Y1111 -

This is a specific case of the following general result. THEOREM 2.1.1.

The interaction effect of factors 91, q 2 , . . . ,qt in a 2k factorial experiment is estimated by 1/2k-' times the difference between the sum of the treatment combinations with

and the sum of the treatment combinations with

+ + . . . + zq, = (t + 1)

zql zq2

(mod 2).

Pmo$ We will prove the result by induction on t. The main effect of factor 41 is the difference between the sum of the treatment combinations receiving factor 41 at the high level (and so with zql = 1) and the sum of the treatment combinations receiving factor q1 at the low level (and so with zql = 0). Thus the result holds f o r t = 1. The interaction between the factors q1 and q 2 is the difference between the main effect of q1 at the high level of q2 and the main effect of q1 at the low level of q 2 . Thus we get


19

0

Hence the result follows by induction.

EXAMPLE 2.1.6. Let k = 4. Then the A13C interaction is estimated by 1/8 of the difference between the sum of the treatment combinations with 2 1 + 2 2 2 3 = 1 (mod 2 ) and the treatment combinations with .rl +n2 +r3 = 0 (mod 2 ) ; this is consistent with the result of Example 2.1 5. 0

+

2.1.2

3'"Designs

In a 3 k design there are k factors and all of the factors have three levels each. The factor might be qunntitntive, such as the tempernture with levels 40°,SO", and 60".or the factor might be qimlitntive, such as severity of symptoms with levels mild, moderate, and severe. I n either case we are again interested i n the effect of each factor, independently, on the response and on the joint effect of two or more factors on the response. 2.1.2.1 Main Effects The main efeerr of a factor is the effect of that factor, independent of any other factor. on the response. So it seems natural, for each factor, to group the 3 k responses to the treatment combinations into three sets. For factor q , these sets are { Y , P ~ T.... ~ r i /.xy = O}, {yI1sz...TAlrq = 1} and {yZ,, *...zi. Inq = 2 ) . When a factor had two levels, we compared the average response in each of the two sets of responses. When there are three sets, there are a number of comparisons that might be of interest. We might be interested in comparing the average response of the three sets pairwise, giving:

which compares responses to levels 1 and 0;

which compares responses to levels 2 and 0; and

which compares responses to levels 2 and I . But these three comparisons are not independent, since we can calculate the third comparison if we know the other two. How can we get a set of independent comparisons'?

20

FACTORIAL DESIGNS

We want each comparison to be independent of the mean of all the responses and of every other comparison so that each comparison can be tested independently of every other N ( p ,C) then a’y N(a’p,a’Ca). comparison and of the mean. (Recall that if y By definition, two normally distributed random variables, x and z are independent if Cov(x,z) = 0. If A is a matrix and if y N ( p ,C ) , then Ay N ( A p ,A C A ’ ) . These results can be used to establish when comparisons are independent; see below and Exercise 2.1.4.10.) We now give some definitions to formalize the ideas that we need. A comparison is any linear combination of the responses. We will write a comparison as

-

N

-

N

Xk

TlX2

&.

Thus the mean is a comparison with ,A, Zk = Two comparisons with coefficients X,, and v,,,, J k are orthogonal or independent if XT1Z2

Ill2

Xk’TITZ

Xk

=’

Zh

(see Exercise 2.1.4.10). Thus a comparison is independent of the mean if 21x2 ...x k

Any such comparison is called a contrast. We want to find two contrasts that are also orthogonal (equivalently pairwise independent). Since all the treatment combinations in each set will have the same coefficient, there are only three distinct coefficients that we need to find. We let XO be the coefficient for all the treatment combinations in the set {y51x2...xk Jxq= 0}, XI be the coefficient for all / z q= l}, and we let X2 be the coefficient the treatment combinations in the set {y5152..,5c for all the treatment combinations i n the set {y51xz...xk/x9 = 2 ) . We define v0,vl and vz similarly. So we need to find XO, XI, XZ, vo,vl and v2 such that XO A1 XZ = 0, vo v1 v2 = 0 and Xovo X l v l X2vz = 0. While there are infinitely many solutions to these equations, there is one solution that is commonly used. Let XO = -1, X i = 0 and A 2 = 1, and let vo = v2 = 1 and v1 = -2. Then these contrasts give the linear and quadratic components of the response to a quantitative factor with equally spaced levels; we prove this below. Any two independent sets of contrasts will give the same sum of squares for a main effect for a normally distributed random vector (see Exercise 2.1.4.10).

+ +

+

+ +

+

EXAMPLE 2.1.7. = 3. Then the three sets of responses for the first factor are

Suppose that k

{ Y O O O ~ YOOI YOOZ, yoio

YO^ 1. yoiz,YOZO. yozi , yo22

1,

~~ioo,~~oi~~ioz~~iio,~iii,~iiz,~izo,~~zi,~izz}, {~zoo,~zoi,~zoz,~21o,~zii,~ziz,~zzo,~zzi,~222}~

The contrast corresponding to Xo = -1, 1 F ( ( Y 2 0 0

-

A1

= 0 and X2 = 1 is

+ Y201 + Y202 + YZlO + y211 + y212 + Y220 + yzz1 + yzzz)

(YOOO

+ yo01 + Yooz + yolo + Yo11 + yoiz + yozo + ~ 0 2 1+ ~022)).


21

21.2.2 Orthogonal Polynomial Contrasts Suppose that we have a set of pairs (x,,y ) ,i = 1 , 2 . . . . . f and that we want to predict Y as a polynomial function of .c of order n = 0.1. . . . f - 1. We usually do this by constructing a matrix X with the first column with entries of 1 (for the constant term), second column with the values of the r ? , the third column with the values of the xp and so on until the final column of X contains the values of 1:;. Thus we can write E ( Y ) = Xp and ,6 = ( X ’ X ) - ’ X ’ Y . In general, nothing can be said about the form of the matrix X’X. However, we could just as readily use other polynomials to determine the entries in the X matrix so that the resulting X ‘ X is diagonal (and hence easy to invert). To do this we proceed as follows. We define ~

for 72 = 1 , 2 , .. . and we let P O ( r )= 1. We will then do a regression on PI (x).P*(.T)and so on, and we want the resulting X matrix to be such that X’X is diagonal. The off-diagonal entries in X ’ X are of the form

and so we require this sum to equal 0. We will assume that the values of the z, are equally spaced; thus, xi+l

- .T,

= rl for all

7.

The values of the coefficients in P,(z) depend on the value of 0. We illustrate the calculations using E = 3. As P = 3, we want to evaluate Po(.), Pl(z)and P2(z). But we already know that Po(.) = 1 (by assumption, since Po(z)is some constant function). Using Equation (2.1) with n, = 1 gives 3

Now substituting for PI we get 3

3

i=l

1=1

So we see that c q , 0 = -y and hence PI (z) = z - F. Using Equation (2.1) with n = 2 and j = 1 gives

=

0.

(2.2)

22

FACTORIAL DESIGNS

Using Equation (2.1) with n = 2 and j = 2 gives

c 3

3

pz(.z)po(.t)

C(xe+

=

t= 1

Q2,152

+m.0)

2=1

0.

=

(2.3)

We now use these two equations to solve for C Y ~ Jand a2,o. From Equation (2.2) we get (collecting terms) a2.1

(Ex:

-3 2 )

( 3 P - 3P)

fa2,o

+Ex:-??Ex: 0.

2

=

I

from which we see that a2.1 =

1

zC2xf- C A

c,x;322 -

.

Using Equation ( 2 3 , we get that i

These are general expressions and we now simplify them by considering the case when x1 = 0 , = ~1 and 5 3 = 2 . Then

+ 1 + 2 ) / 3 = 1, = 0 + 1 + 4 = 5.

z = (0

t

and =0

+ 1 + 8 = 9.

2

So we get a2,1 =

(I x 5

-

9)/(5 - 3 x 1) = -2

and ~12.0

= (-5 - 3 x ( - 2 ) x 1 ) / 3 = 1 / 3 .

We usually record the values of the orthogonal polynomials for a given value of n rather than the actual polynomials. Hence we would get a table like Table 2.1 for n = 3. Table 2.1 Values of Orthogonal Polynomials for n = 3 X

PI(.) 3P~(s)

0

1

2

-1 1

0 -2

1 1

Linearcomponent Quadratic component

We can find the values of orthogonal polynomials for any value of n in a similar way. Tables of orthogonal polynomials may be found in Kuehl(1999) and Montgomery (2001). It is possible to evaluate orthogonal polynomials for unequally spaced levels; see Addelman (1 962) and Narula (1 978). For a factor with !. levels, the values of (z) are the coefficients of the linear contrast and the values of Pz (z)are the coefficients of the quadratic contrast. Higher-ordercontrasts are defined similarly.


23

2.1.2.3 interaction Effects We start with two-factor interactions. The two-factor interaction effect of factors A and B is the joint effect of the two factors on the response and is again called AB. Now, however, there are four independent contrasts associated with the interaction. This is because there are nine distinct pairs of levels for factors R and B and so 8 independent contrasts possible between these sets. Four of these contrasts have been used to calculate the main effects of the factors, leaving 4 contrasts for the interaction effect. The easiest way to determine these contrasts is to divide the responses to the treatment combinations into nine sets, basedjointly on the levels of A and B. So we get sets

ARou

=

for 0 = 0 , l . 2 : 4') = 0.1.2; lzn = Q , x g = I,!!},

(yZlTZ

for any two factors A and B. There are 8 independent contrasts possible between 9 sets. However, these sets are subsets of the sets that we used to define the main effects of factors A and R. To ensure that the main effects and the interaction effect are independent, we have to use those four contrasts (two from each main effect) to be 4 of the X contrasts. Then any other four independent contrasts can be chosen to represent the interaction effect. One common way to get the final four independent contrasts is to take the component-wise product of the contrasts for main effects. This gives contrasts that are interpretable as the linear x linear interaction, the linear x quadratic interaction, the quadra/ic x linear interaction and the quadratic x quadratic interaction. It is easy to verify that these contrasts are independent of the contrasts for main effects. W EXAMPLE 2.1.8. Suppose that k = 2. 'Then the contrasts for main effects and for the linear x linear contrast, the linear x quadratic contrast, the quadratic x linear contrast and the quadratic x quadratic contrast are given in Table 2.2. 0 Table 2.2 A . B, and A B Contrasts for a 32 Factorial 00

II

01

02

10

1

1

1

-1 1 0 -2 0 2 0 -2

I -1

I

-1 I -1 I I -1 -1 I

0

0

I I I -1 -1 1 1

-2 -1 1

0 0 2 -2

-2 0 -2 0 0 0

4

12

20

21

1

1

0 -2

1

1 1 1

1

I 0 0 -2 -2

1

-1 I -1 1 -1 1

0 -2 0 -2 0 -2

22

Treatmentcombinations

IMean I ALinear = A L 1 AQuadratic = A Q 1 BLinear = B L I BQuadratic = BQ I ALXBL 1 ALXBQ I AQXBL I AQXBQ

Another way to get the four contrasts that correspond to the two-factor interaction effects is to use the sets {YX1T2. zi.lzql + x q 2 = Q}? 8 = 0.1.2 and {YrT1.rz...?l'A lzgl 22q2 = Q}: 8 = 0 . 1 . 2 .

+

Then the linear and quadratic contrasts can be used on these sets, although it is not clear what interpretation might be put on them, even for quantitative factors. Note, though, that

24

FACTORIAL DESIGNS

these sets are the pencils of an affine plane and that this link has been exploited in proving results about fractional factorial designs; see Section 2.2 and Raghavarao (1971) for more details. EXAMPLE 2.1.9. Suppose that k = 3. The set given by

z1

+ z~ = 0 is

If we let

ABoo

= ( 2 1 ~ 2 ~ 3 1 2= 1 O , X ~= 0 } ,

ABol

= ( 2 1 2 2 2 3 1 2 1 = O , 2 2 = I},

+

and so on, then the set given by 5 1 ICZ = 0 is the union of ABoo, AB12,and ABzl. The other 5 sets can be defined similarly. The 9 orthogonal contrasts are then as given in 0 Table 2.3. Table 2.3 A. B , and A B Contrasts for a 33 Factorial ABoo

ABol

ABoz

ABio

ABii

I 0 -2 -1 1 0

-20

1

1

-1 1 -1 1 -1

-1 1 0 -2 0

I -1 1 1 1 1

1

-2 1 1

1 0 -2

-1 1

-2

AB12

ABzo

1

1

0

0

1 1

1 1

-2 0 -2 1

-2 1 1 -1

1 -1 1 1

1 0 -2 -1

-11

1

1

1

1

1

AB21

-2

AB22 I

Mean

:i AEffect

1 01

I;}

B Effect Al? Effect

1

The definition of the sets of treatment combinations for the determination of higher-order interaction terms is similar. We can either take the orthogonal polynomial approach with linear x linear x linear and so on, or we can use the second, “geometric”, approach. With the geometric approach, the three-factor interaction between factors A , B , and C requires four sets to define the eight contrasts; these are

xi

+ 252 + 2 9 = i

(mod 3),

z1

+ 2x2 + 2x3 = i

(mod 3 ) .

We will take the geometric approach when we come to construct fractional factorial designs in Section 2.2. 2.1.3

Asymmetric Designs

As we said earlier, we use x Cz x . . . x e k to refer to a general factorial design in which there are no restrictions on the number of levels for any of the factors.

25


2.7.3.1 Main Effects As before, we want to be able to estimate the effect of each of the factors, individually, on the response. For factor q , we calculate Fq sets associated with that factor and use contrasts between these sets to calculate the main effects for that factor. EXAMPLE 2.1.10.

Consider a 2 x 3 x 4 complete factorial design. The two sets for the first factor are

{000.001,002.003.010.011,012.013~020.021~022~023}~ {100.101,102,10~3.110,111,112,113,120.121,122,123).

The second factor has three sets associated with it: (000,001.002.003, LOO, 101.102,103}. {010,011~012~013.110.111.112.113}, {020.021.022.023,120,121,122.123}.

Similarly, there are four sets associated with the third factor. The corresponding matrix of contrasts for main effects for the three factors is given in Table 2.4. 0 Table 2.4 Main Effects Contrasts for a 2 x 3 x 4 Factorial

-1 - 1 I

-1 -1 I

-'1

-1

I

-1

-1

1

-1 -1 1

I

-1 -3

--1 - 1 -1 -1 -1 --I 11 n 11 11 1 - 2 - 2 - 2 - 2 I -3 -1 I 3 I I -1 -1 1 I -1 3 --J I

-1

-1

I I

I

4 -1 1 -1

-1 I I

I

I

-1 I

-1 -3

-1 I

I --I

I

I

-1

I

I

-1 I

3

-3

-1

1 I

I -1

- 1 I

I I I I! 11 11 1 - 2 - 2 - 2 - 2 3 --J -1 I I I -1 -1 I -1 3 -3 I

-1

I I

-1 -3

I

11

I 1 I

I

I

I

I I

I

I

A iiL

I l i l ( 2

1 - J - 1 I l ( ' i . I I -I -I ('(2 1 - 1 l - : < l < ' c

,

2.1.3.2 Interaction Effects For interaction effects, we again use the polynomial contrasts for each factor and take all possible component-wise products. EXAMPLE 2.1.11.

Consider the 2 x 3 factorial. Then the six treatment combinations are 00, 01, 02, 10, 1 I , and 12. The contrast for the main effect of the first factor is -1, -1, -1, 1, I , I . Forthesecondfactorthemain-effectcontrastsare-l,O, I , -1,0, 1 and I , -2, I , 1, -2, 1 . Thus the two contrasts for the interaction are 1, 0, -1, -1, 0, 1 and -1, 2 .-I, I , -2, I .

0 2.1.4

Exercises

1 . Give all the level combinations in a 2 x 2 x 3 factorial 2 . Give the B and C: main effects for Example 2.1.3. 3. Use m(Ez,j=l y51r2...IA - C,,,=o y I l T 2....T i ) to confirm the results in Example 2.1.3 and the previous exercise. 1

26

FACTORIAL DESIGNS

4. Consider factors with two levels and let k = 2 Confirm that the AB interaction and the B A interaction are equal. Repeat with k = 3. 5. Consider factors with two levels and let k = 3. Show that the ABC interaction is also equal to the average difference between the AC interaction at level 1 of B and the AC interaction at level 0 of B . Are there any other ways that you could describe this interaction? 6. If .t = 4, verify that the following components are correct -3 -1 1 3 Linear 1 -1 -1 1 Quadratic -1 3 -3 1 Cubic

7. Suppose that k = 2 and that both attributes have 4 levels. Give possible entries for the 6 polynomial contrasts for main effects and the 9 polynomial contrasts for the two-factor interaction. 8. Suppose that .t = 3.

(a) (b) (c) (d) 9.

Let x = (0,1,2)’. Let X = [I,x.x2].Calculate X’X. Now let X = [l,PI ( x )P~z ( z ) ]and calculate X’X. Comment. Suppose that you use the representations in both of the first two parts to fit a quadratic polynomial. What is the relationship between the estimates?

(a) List the 6 sets of treatment combinations in a 3’ factorial experiment corresponding to the main effects. (b) Give two independent contrasts for the main effects for each of the factors. (c) List the six sets that correspond to the two-factor interaction. (d) Give four independent contrasts corresponding to the two-factor interaction. (e) Verify that all the eight contrasts (for main effects and the two-factor interaction) that you get are orthogonal (that is, independent) in this case.

10. Suppose that y1, yz, . . . , yn are independently identically distributed N ( p ,g 2 ) random variables.

(a) Let y’ = ( y l , yz, . . . , yn) and let a’ = ( a l ,az, . . . , a n ) . Then i

xi

Hence deduce that the sum of squares for testing HO : p a, = 0 is given by ( p C ,a1)2n/ts2 where ,s2 is an unbiased estimate of CT’. (b) By definition, two normally distributed random variables, z and z are independent if C O l l ( Z , z ) = 0. If A is a matrix, then Ay N ( A p ,(r2AA’),where p = ( p , p , . . . , p ) . Suppose that the rows of A form a set of independent contrasts. Then AA’ is a diagonal matrix. Derive the sum of squares for testing Ho : Ap. = 0. (c) Hence show that, in an ordinary least squares model, the sum of squares for a factor is independent of which set of independent contrasts is chosen.

-

1 1 . Verify that the five contrasts in Example 2.1.1 1 are all mutually orthogonal.

REGULAR FRACTIONAL FACTORIAL DESIGNS

2.2

27


A froctional factorial design is one in which only a subset of the level combinations appears. Fractional factorial designs are used when the number of treatment combinations i n the complete factorial is just too large to be practical, either because it will take too long to complete the experiment or it will cost too much. So a fractional factorial design is faster and cheaper and, with a suitably chosen fraction, you can still get all the information that you want from the experiment. In this section we will look at the ways of getting fractional factorials that allow you to estimate all the effects you are interested in.

2.2.1

Two-Level Fractions

Recall that, if each of the k factors has 2 levels, then we talk about a 2k design. We use 2"-" to denote a fractional factorial design in which only 2"-P treatment combinations appear. A regirlar,froction of a 'Lk factorial is a fraction in which the treatment combinations can be described by the solution to a set of binary equations. Equivalently, a regular fraction is one in which there are some generator factors and all other factors can be defined in terms of these generators. The binary equations are called the dejning equations or the dejnirzg contrasts of the fractional factorial design. A regular 2k--1, fraction is defined by p independent binary equations or, equivalently, by k - p generators. An irregular fraction is a subset of the treatment combinations from the complete factorial, but the subset is determined in some ad-hoc fashion. Usually all treatments combinations in an irregular fraction have to be listed explicitly. EXAMPLE 2.2.1.

A regular 2"-' has 8 treatment combinations and these 8 treatment combinations are the solutions to one binary equation. The solutions to the binary equation x1 +x2 + x 3 = 0

(mod 2)

+

are the treatment combinations in Table 2.5. If x1 = 1 and 2 2 = 0 then 1 0 + 2 3 = 0 (mod 2); so 2 3 = 1. The defining equation places no restrictions on x4; so both 1010 and 1011 are solutions to the defining equation. This fraction may also be written as I = ARC. (Recall that. when working modulo 2, 0 0 = 1 + 1 = 0 and 0 1 = 1 0 = 1.) 0

+

Table 2.5

+

+

A Regular 24-1 Design

0 0 0 0 1 1 1 1

0 0 0 1 1

0 1 1 0 0 1 1

1 1 0 0

0 1 0 1 0 1 0 1

As we have said, for a regular 2"-P design, the treatment combinations will satisfy p independent binary equations. A set of equations is said to be independent if no non-zero linear combination of the equations is identically 0. We now consider an example of this.

28

FACTORIAL DESIGNS

mEXAMPLE2.2.2. Let k = 6. Consider the three binary equations

To show that these equations are independent, we need to consider all the sums of these equations; thus we consider the three sums of two equations and the sum of all three equations. We get

and these are the only linear combinations possible for binary equations. There are 8 solutions to theseequations: 000000,001l11,010101,011010, 100110, 101001, 110011, I I 1 100. On the other hand, the equations 21 5 2 2 3 2 4 = 0 , q x2 25 2 6 = 0 and 23 24 25 2 6 = 0 have 16 solutions since the third equation is the sum of the first two and so there are only two independent equations. 0

+ + +

+ + +

+ + +

We have defined regular fractions by setting linear combinations of the x2 equal to 0. This is called the principalfraction. Other fractions are obtained by equating some (or all) of the linear combinations to 1 . Indeed if we use the same linear combinations with all possible solutions we obtain a partition of the complete factorial into regular fractions, as the following example shows. HEXAMPLE 2.2.3. Let k = 5 and consider the linear combinations 2

+ +

+ +

1 22 23 and 53 24 z5.The corresponding partition of the complete 25 factorial into four regular fractions is given in Table 2.6.

Table 2.6 Non-overlapping Regular 25-2 Designs

0 0 0 0 0 0 0 0 1 1 0 0 1 1

1 1 0 0

1 1 1 1

0 1 0 1

1 0 1 0

I 1 0 0 0 1 1 0 1 1

0 0 0 0 1 1

0 0 1 1 0 0

0 0 1 1 1 1

0 1 0 1 0 1

1 0 0 1 0 1

I 1 0 0 1 1 1 0 1 0

0 0 0 0 1

1 1 0 0 1

0 0 1 1 1

0 1 0 1 0

0 1 1 0 1

I l l l O 1 0 0 0 0 1 0 0 1 1

0 0 0 0 1 1 1 1

1 1 0 0 1 1 0 0

0 0 1 1 1 1 0 0

0 1 0 1 0 1 0 1

1

0 0 1

0 1

I 0


29

For any fractional factorial design, we will estimate an effect by using the same coefficients as we would have used for that effect i n the complete factorial design. This is sometimes called definition by restriction. wEXAMPLE2.2.4.

Let k = 4 and consider the 24-' fraction given by .TI + 2 2 + 1 3 + 5 4 = 0; see Table 2.7. There are 8 treatment combinations and so there will be 8-1=7 orthogonal contrasts possible. The IS contrasts from a 2" design are given in Table 2.8. The contrasts that we use for the fraction are obtained from the contrasts in Table 2.8 by considering only the eight columns that correspond to the eight treatments in the fractional factorial design. These are indicated in bold. Doing this, we see that A B C D is indistinguishable from the mean. This is consistent with Theorem 2.1.1 since all the treatment combinations with 21 n'2 x3 2-4 = 0 have the same sign i n the A B C D contrast. All the other contrasts come i n pairs with, say, the effects of A and B C D being indistinguishable. This is because the two sets with different coefficients in the A contrast are given by 2'1 = 0 and 21 = 1. For the B C D contrast these sets are 2 2 2 3 2 4 = 1 and 2 2 2 3 + 2 4 = 0. If we know that 11'1 n ' ~ 5 3 2 4 = 0, then 2 1 = x2 + 3 3 + 2 4 and so the sets for the A contrast and the B C D contrast coincide. Similarly, x 2 = x1 x 3 z4, .x3 = x L + xz + .c4, 5 4 = .TI 3:2 2 2 , 21 22 =23 ~ 4 x1 , + .r3 = r z + 2 4 , and z1 + 5 4 = 2 9- + 7'3 So A and HCD, B and ACD, C and ,4BD, D and A B C , AB and C D , A(' and B D , and AD and BC form pairs of effects with contrasts that are the same in the 2"' fraction given by x1 1 2 , ~ $3 x$ = 0.

+ + +

+ + + + + +

+ +

+ + +

+

% I . .

+ +

Table 2.7 A 24p1 Design of Resolution 4 0 0 0 0 0 0 1 I 0 0 1 1 1

1 1 0 0 1

0 1 0 1 0

1 0 1 0 0

I 1 1 1

We say that effects with the same contrast i n a fraction are aliased or confomdedeffects. The effects which have the same coefficient for every treatment combination in the fraction are called the dejning efferts or the dejning contrasts. The list of all the aliases for each effect in a design is called the alias .strurture of the design. Given the defining contrasts, we can calculate the alias structure (and conversely). I f we know nothing about the factors and possible interactions, then it is usually best to be able to estimate low order interactions (main effects and two-factor interactions) independently of each other i n a fractional factorial design. So fractional factorial designs are classified by the alias structure of the design. If no main effect is confounded with any other main effect, but at least one main effect is confounded with a two-factor interaction, then the design is said to be of resolution 3. If at least one main effect is confounded with a three-factor interaction but no two main effects are confounded with each other and no main effect is confounded with a two-factor interaction, then the design is said to be of resolution 4. So the design in Table 2.7 is of resolution 4 since each main effect is

30

FACTORIAL DESIGNS

Table 2.8 Contrasts for the 24 design 0 0 0 0

A B C D AB AC AD BC BD CD ABC ABD ACD BCD ABCD

0 0 0 1

1 1 1 0

1 1 1 1

-1 -1 -1 -1 -1 -1 -1 -1 1 1 1 1 1 1 I -1 -1 -1 -1 1 1 1 1 -1 -1 -1 -1 1 1 I -1 -1 1 1 -1 -1 1 1 -1 -1 1 I -1 -1 1 -1 1 -1 1 -1 1 -1 1 -1 1 -1 I -1 1 -1 1 1 1 1 1 -1 -1 -1 -1 -1 -1 -1 -1 1 1 1 1 -1 -1 1 1 1 -1 -1 -1 -1 1 1 -1 -1

1 1 1 1 1 1 1 1

1 1 1 1 -1 -1 -1 -1

1

0 0 1 0

0 0 1 1

0 1 0 0

-1 1 -1 1 -1 -1 -1 1 -1 -1 -1 1 1 1 -1 I -1 1 1 -1 1 1 -1 1 -1 -1 1

0 1 0 1

0 1 1 0

0 1 1 1

1 0 0 0

1 0 0 1

1 0 1 0

1 0 1 1

1 1 0 0

1 -1 1 -1 -1 1 -1 1 1 1 1 1 -1 -1 -1 -1 -1 1 -1 I 1 -1 1 -1 1 -1 -1 1 1 -1 -1 I 1 1 -1 -1 1 1 -1 -1 1 -1 1 -1 1 -1 1 -1 -1 1 1 -1 1 -1 -1 1 1 1 -1 1 -1 -1 1 -1 -1 1 1 -1 -1 1 1 -1

1 1 0 1

-1 1 -1 -1 -1 1 1 -1 -1 -1 -1 1 1 -1 1 -1 1 -1

-1 1 -1

-1 1 -1 -1 -1 -1

1 1 1 1 1 1 1

confounded with a three-factor interaction. If at least one main effect is confounded with a four-factor interaction and no main effect is confounded with anything smaller, or if at least one two-factor interaction is confounded with a three-factor interaction but no pair of two-factor interactions are confounded, then the design is said to be of resolution 5 . We can determine the resolution of a design directly from the defining equations. If there are r non-zero coefficients in a defining equation, then main effects corresponding to the zqwith non-zero coefficients in the equation are confounded with interactions of r - 1 factors, two-factor interactions corresponding to pairs of zq with non-zero coefficients are confounded with interactions with r - 2 factors, and so on. EXAMPLE 2.2.5.

Let k = 5 and consider the fraction given by the defining equations 21

+ 22 + 2 3 = 0 and + + + 25 = 0. 22

24

The aliasing structure is determined by noting, first, that 21

+ 2 2 + +2 1 + +5 4 +

+24 +

2 5

= 0.

+ + 2 3 = + 2 2 + 2 4 + 5 5 = .23 + 2 4 +

25

= 0.

23

22

25

= 53

Thus there are in total three defining equations; these are 21

22

21

From these we can now establish the sets of equations that have the same solutions. We get 51

52 23

24 25

= = = = =

-

22 +x3

.c1 + x 3 51

21

x1 + x 2 +a2 2 3

+ + + + 53 + 22

2 4 2 5

= = =

2 1

+ + + + x5 + + + + + x2 + 2s + + 22

x4

21

x4

22

x3

24

22

54

2 5

21

51

2.5

= = =

-

2 1

52

+ + T4 + + + + + xs. + + 23

53

x4

24

23

3-5,

x3

24

55, x5,


31

Thus we see that the main effect of A , for instance, is confounded with BC, B D E and 4CDE. As there is at least one main effect that is confounded with a two-factor interaction (and no pair of main effects that are confounded), we have confinned that the design is of resolution 3. 0 EXAMPLE 2.2.6. Let k = 6. The defining equations 3'1

+ 5 2 + ~3 + 2 4 = 0 and 2 3 + 2 4 + 2 , + 3'6 = 0

give a design of resolution 4. We can see this because 3'1

+ .x2 + 2 2

t 24

+ ~3 + 3'4 + + 2 6 = 1-1 + .r2 + 3'5 + .rc = 0

Thus each main effect is confounded with two 3-factor interactions and a S-factor interaction. Each 2-factor interaction is confounded with at least one other 2-factor interaction. If there are more than three factors ( k > 3), then we can find a resolution 3 design that uses fewer treatment combinations than the complete factorial. Designs of resolution S can be used to estimate main effects and two-factor interactions. If there are four or fewer factors (so k 5 4), then the only resolution 5 design is the complete factorial, independent of the number of levels each of the factors have. If there are more than four factors ( k > 4), then we can find a resolution S design that uses fewer treatment combinations than the complete factorial. We do this by finding a set of defining equations such that each defining equation has at least S non-zero coefficients and such that each linear combination of the equations has at least S non-zero coefficients just as we did for resolution 4 in Example 2.2.6.

2.2.7.7 Regular Designs with Factors at 2 Levels Regular fractions can be constructed from a set of defining equations or, equivalently, from a set of generator vectors. In this section we give sets of generator vectors that can be used to construct small regular fractional factorial designs with all factors with 2 levels. For two vectors a and b we will define a b using component-wise addition modulo 2 (that is, 0 0 = 1 + 1 = 0 and 0 1 = 1 + 0 = 1). Thus for the two vectors

+

+

+

a = ( 0 . 0 . 0 . 0 . 1 , 1 , 1 . 1 ) and b = (0.0.1.1.0,0.1.1)

we see that a+ b = (0~0,1~1,1~1,0,0).

We will write a set of gariernror vertors i n a standard order. If the design has 2 k - p treatments then we need to define k - p generators, bi say. We let bl have its first 2"-"-' entries as 0 and the remaining 2"P-l entries as 1. For b2 we let the first 2"-"-2 entries be 0, the next 2"p-2 entries be I , the next 2k--P-2 be 0 and the final 2 k - p - 2 be 1. For b:3 we let the first 2k-p-3 entries be 0, the next 2k--P-3 entries be 1 , the next 2 k - p - 3 be 0 and so on. We continue defining generators i n this way until we get to b k P p which alternates 0s and Is. For instance when k == 4 and p = 1 we get the three generator vectors bl = ( 0 , 0 . 0 . 0 . 1 , 1 . l . l ) . b *= ( 0 , 0 ~ 1 ~ 1 , 0 , 01) , 1 ,and

b3

= (0.1.0.1.0.1.0.1).

We can now define the levels of all the other factors in the design in terms of the b,. Table 2.9 contains designs with up to 10 factors of resolution 3 and Table 2.10 contains

32

FACTORIAL DESIGNS

Table 2.9 Smallest Known 2-Level Designs with Resolution at Least 3

k

N

Other Factors

4

8

5

8

6

I

8 8

8

16

9

16

10

16

b4 = bl b2 b3 b4 = bi bz. b 5 = bl b3 b4 = bl bz, b5 = bi b3, bg = bz b3 b4=bi+bz,b5=bi+b3,bs=bz+b3, b7 = bl bz b3 b5 = bz b3 b4,bg = bi b3 b4, b7 = bl bz+ b3,bs = bi bz b4 b5=bz+b3+b4,b6=bl+b3+b4, b7 = bi b2 b3. b8 = bi b2 b4, bg = bl bz b3 b4 b5 = b2 +b3 b4,bs = bi f b 3 b4. b7 = bl b2 b3, b8 = bi bz b4, bg = bi +b2 +b3 bq,bio = bi +bz

+

+

+ +

+

+

+ + + + + + + + + + + + + + + + + +

+

+ + +

+ +

Table 2.10 Smallest Known 2-Level Designs with Resolution at Least S k

N

Other Factors

5 6 7 8

16 32 64 64

9

128

10

128

ba = bl bz b3 bq be = bl bz b3 + b4 + bs b7 = bl bz b3 bq b5 + bg b7 = bi b2 b3 b4. bs = bi bz b5 + bg b8 = bl +b3 b4 + b 6 + b7, b9 = bz b3 bg btj + b7 bs = bl bz b3 -tb7, b9 = bz b3 b4 b5, blo = bi b3 b4 b.5

+ + + + + + + + + + + + + + + + + + + + + + + + + +


designs with up to 10 factors of resolution 5 . We let combinations so IV = 2"p.

33

denote the number of treatment

EXAMPLE2.2.7.

We use the results in Table 2.9 to construct a binary design with k = 6 factors of resolution 3. The 8 runs of the design are given in Table 2.1 I . 0

Table 2.11

A Design with 6 Binary Factor? of Resolution 3

0 0 O

0 0 l

0 1 1 1 1

1 0 0 1 1

0

0

0

0

1 O 1 0 1 0 1

0 l 1 1 1 0 0

1 O 1 1 0 1 0

1 l 0 0 1 1 0

2.2.2 Three-Level Fractions A regular 3 k - p fraction is one that is defined by the solutions to a set of p independent ternary equations; that is, a set of equations where all the arithmetic is done modulo 3 (so 0 + 2 = 1 + 1 = 2 + 0 = 2 , 1 + 2 = 2 + 1 = O a n d O + l = 1 t O = 2 + 2 = 1). These equations are the defining equations or the defining rontrusts of the regular fractional factorial design. EXAMPLE 2.2.8.

A regular 3'-' fraction has 9 treatment combinations which are the solutions to two independent ternary equations. The solutions to 2 1 2 2 2 3 = 0 and 21 2x2 ~4 = 0 are given in Table 2.12. (If we had used 2 1 2 2 2 3 = 0 and x1 52 x4 = 0, then we would have had 2 3 = x4,and so we would not have been able to estimate the effects of the third and fourth factors independently.) 0

+ +

+ +

+ +

+

+

A regular 3 k - p fraction satisfies p independent ternary equations. When checking for independence, we must now add the original equations in pairs, triples, and so on (as you do for binary equations), but we must also check each equation plus twice every other equation, and so on. More formally, if we let El E z : . . . . E p be the defining contrasts, then we must calculate ~

and show that none of these 3 p equations are identically 0 except when CYI

= CY2 =

. . . = c l p = 0.

34

FACTORIAL DESIGNS

Table 2.12 A 34-2 Fractional Factorial Design 0 0 0 1 1 1 2 2 2

0 1 2 0 1 2 0 1 2

0 2 1 2 1 0 1 0 2

0 1 2 2 0 1 1 2 0

HEXAMPLE 2.2.9. Let k = 6. Consider the defining contrasts 51

+ 5 2 + 2 3 + 2 2 4 = 0, 2 1 + 2 2 + 2 2 3 + 2z5 = 0

These are all independent since each of 2 4 ,

55

and

2 1

+ 2 2 2 + x3 + 2z6 = 0.

and 2 6 is involved in only one equation.

0

EXAMPLE 2.2.10.

Let k = 5 . Consider the defining contrasts 21

+22 +2 3 = 0

and

23

+ 2 4 + 25 = 0.

Can we find a third independent equation with at least three non-zero coefficients? The two equations that are linear combinations of the defining contrasts are 21

+ 2 2 + 2 3 + 5 3 + 2 4 + x 5 = z1 + 2 2

+ 223

z1

+ + +2x3 +

52

and

22

23

2x4

+

54

+

25

=

+ 2x5 = + + 2 2 4 + 2x5 = 0. 21

(We do not need to double the first equation and add the second equation - this is just double the sum of the first equation and twice the second equation; similarly, if we double both equations and add, we just get double the sum of the two equations.) Any equation other than the original two equations and the two equations we have found by addition will be independent of the two equations. What can we say about the possible equations? Given we are using the equations to construct fractional factorial designs, we do not want an equation with only one zq in it since the corresponding factor would not vary in the experiment. Similarly, if we use zql m q lthen , there is a constant relationship between the corresponding factors and the associated main effects would be aliased. If we think about equations with three variables, then we cannot have more than one of 5 1 , 2 2 and 2 3 (since if we used, say, z1 22 5 4 = 0, then

+

+ + 2 1 + 2 2 + 2 3 -k 2 ( 2 1 + 2 2 + 2 4 ) = 2 3 + 2 2 4 = 0

and so 2 3 = 2 4 ) and we can not use more than one of 2 3 , involving three of the zq forces two of the zq to be equal.

54

and

25.

So an equation


35

A similar argument shows that you can not have equations with four or five variables either. 0

Since any interaction sum of squares is independent of the particular contrasts that are used to define the interaction effects, in this section we will define the interactions using contrasts between sets ol'the form

The only attributes whose levels determine the entries i n the sets are those with non-zero coefficients. There are three sets associated with each equation: {(.rl.,Q..

. . ,x k ) l

C o q x q= 01,

{ ( x l , r Z , .. . , :rk)l

C oqxq = I 1 9

'I

We will let

P(&z,)

=

P ( n 1 . 0 2 , .. . . O h )

ri

C,

be the set of three sets associated with the equation oq.rrl, and we will talk about the pencil associated with C , o q ; r q . For any interaction involving f attributes, there are 2t-' associated pencils. For example, if we want to calculate the interaction between the first three factors we would calculate contrasts between the sets i n P ( 1110.. . O), P ( 112 0 . . . O), P ( I21 0 . . . O ) and P ( 1220. . . O ) . Each pencil gives rise to two independent contrasts. Contrasts from different pencils are orthogonal since any two sets from different pencils intersect in 3'-' treatment combinations. EXAMPLE 2.2.11.

Let X: = 2. The sets i n each pencil are indicated in Table 2.13. The corresponding contrasts are given in Table 2.14. 0 Table 2.13

Pencil P(10) P(O1) P(11) P( 12)

or

TI

or

~2

or or

z1+r2=6' TI

= 6' = 6'

+ 2rl

=

6'

Pencils for a 3' Factorial Design 6'=0

6'=1

0=2

00,01.02 00, 10. 20 00, 12, 21 00. 1 I , 22

10. I I . 12 01. I I . 21 01. 10.22 02, 10.21

20,21.22 02, 12. 22 02, I I . 20 01. 12. 20

EXAMPLE 2.2.12.

Consider the design constructed in Example 2.2.8 and given i n Table 2.12. The contraqts for the main effects for the design are given in Table 2 . IS.

36

FACTORIAL DESIGNS

Table 2.14 Contrasts for a 3' Factorial Design 00 01 02 -1-1-1 1 1 -1 0 1-2 -1 0 1-2 -1 1 1 1 - 2

1 1 1 I 1 0 -

10

II

12 20 21

0 0 0 1 I -2-2-2 1 1 -1 0 1-1 0 1-2 1 1-2 0 1-1 1-1 -2 1 1 I 10 - 1 1 1 02 I 1 1-1

22

I 1 1 1

0 2 1 1

Table 2.15 Contrasts for a S4-' Factorial Design 0000 0121 0212 1022 I l l 0 1201 2011 2102 2220 - 1 - 1 - 1 0 0 0 1 1 1 - 2 - 2 - 2 -1 0 1 - 1 0 1 1 - 2 1 1 - 2 1 1 0 1 0 - 1 -1 1 1 - 2 1 - 2 1 -1 0 1 1 - 1 0 1 - 2 1 1 1 - 2 -

I 1 I I 1 I - 1 0 1 1 - 2 I 0 - 1 1 - 2 1 1 0 1 - 1 2 I 1

AL AQ

BL BQ CL

CQ DL DQ


37

2.2.2.7 Regular Designs with All Factors at 3 Levels For two vectors a and b, we will define a b using component-wise addition modulo 3. Thus

+

0+0= 1 + 2 = 2 + 1 = 0, 0 + 1 = 1 + 0 = 2 + 2 = 1 and 0 + 2 = 1 + 1 = 2 + 0 = 2 . Hence, for the two vectors a = (0,O.0,1, 1 . 1 . 2 . 2 . 2 ) and b = (0.1,2.0: 1 . 2 . 0 , l 2), ~ we see t h a t a + b = (0. 1 . 2 . 1 , 2 , 0 , 2 . 0 , 1). We will write a set of genernfors in a standard order. If the design has s k - p treatments, then we need to define k - p generator vectors b,. For bl, we let the first 3"-"-' entries be 0, the next 3 k - p - ' entries be 1 and the remaining 3"-P-' entries be 2. For b?, we let the first 3k-71-2 entries be 0, the next 3"-P-' entries be I , the next 3"-P-' be 2, the next 3"-"-? entries be 0, and so on. For b3, we let the first 3"-PP3 entries be 0, the next 3k-p-3 entries be I , the next 3k-1)-3 be 2, and so on. We continue defining generators in this way until we get to b k - p which has 0, 1 and 2 in turn. For instance, when k = 4 and p = 1, we get the three generator vectors bl = (O.O,O.O.O,0.0,0,0, 1 . 1 , 1 , 1 :1.1.1: 1,1.2.2.2.2.2.2.2.2,2),

bz

=

(0.0.0.1.1.1.2.2.2,0.0,0.1,1,1,2.2.2.0.0.0.1.1.1.2.2.2).

and b3 =

( 0 . 1 . 2 . 0 . 1 . 2 . 0 , l . 2,0,I , 2 . 0 . 1 , 2 .0.1.2.0.1,2,0.1.2.0. I , 2).

If we write 2a, then that means we multiply each component of a by 2 modulo 3 We can now define the levels of all the other factors i n the design i n terms of the b,. Table 2.16 contains generators for designs of resolution 3 with up to 10 factors and Table 2.1 7 contains generators for designs of resolution 5 with up to 10 factors We let AT denote the number of combinations of factor levels so N = 3--P. Table 2.16 Smallest Known Regular ?,-Level Designs with Resolution at Least 3 k

N

4

5 6 7

9 21 21 21

8

27

9

27

10

21

2.2.3

A Brief Introduction to Finite Fields

The ideas developed in the previous two sections make it easy to construct fractions that confound certain effects. The satne ideas will not work for all possible numbers of levels

38

FACTORIAL DESIGNS

Table 2.17 Smallest Known 3-Level Designs with Resolution at Least 5 k 5 6 I

N 81

8

243 243 243

9

243

10

243

nor for situations in which not all factors have the same number of levels. In this section we develop the idea of aJiniteJield. This is an algebraic structure that we need in order to be able to extend the techniques of the previous two sections. We hegin by looking at four levels. Suppose that there are k = 2 factors. Suppose that we work modulo 4 so that 1 1 = 2, 1 2 = 3, 1 3 = 0 , 2 2 = 0 , 2 + 3 = 1, and so on. Think about the pencil P(12). It contains the four sets

+

+

{(21,22)121

+

+

+ 2 ~ 2= 0 ) = {00,02,21,23),

{ ( ~ l , z 2 ) l z+222 l

= 1) = {10,12,31,33},

{ ( ~ 1 , ~ 2 ) 1+ ~21 ~

2= 2) = {01,03,20,22},

{ ( 2 1 , 2 2 ) 1 ~ 1+

2 ~ 2= 3) = {11,13,30,32).

Note that each of these sets contains either two entries or zero entries from the sets of the pencil P( 10) (from which we calculate the A main effect). Since these intersections are not of a constant size, the arguments that we have used in the previous sections, to establish the orthogonality of the contrasts used to estimate the effects, do not work. How can we overcome this problem? The reason that all the sets from different pencils intersected in a constant number of points when working modulo 2 and 3 is that every non-zero element could be multiplied by some other element to get 1; that is, all non-zero elements have multiplicative inverses. For instance, 1 x 1 = 1 modulo both 2 and 3 and 2 x 2 = 1 modulo 3. When we work modulo 4 we see that 2 x 1 = 2, 2 x 2 = 0 and 2 x 3 = 2. Thus 2 does not have a multiplicative inverse. So we want to get a set of 4 elements in which multiplicative inverses exist. We do this by using an appropriate polynomial which we evaluate over the integers modulo 2 , 2 2 . To get such a polynomial, consider the quadratics over 2 2 , namely, x 2 , x 2 z = x(z l ) ,x 2 1 = (x 1)2and z2 z + 1. The first three quadratic equations factor over 2 2 , but l 2 1 1 = 1 and 0’ + 0 + 1 = 1 in Zz, and hence the quadratic polynomial z2 z 1 does not factor over Z2. It is said to be irreducible over 2 2 . We now try to embed 2 2 in a larger field in which x2 z + 1 will factor. Suppose we let cy be a solution of z2 z + 1 = 0. So a2 N 1 = 0 and hence o2 = N + 1. Since we are working modulo 2 , we have

+

+ + + + + + +

+

+

+

+ +

( a + 1 ) 2 + ( C Y + I ) + l = ( c y 2 + 1 ) + ( c y + 1 ) + 1 = a 2 + o + 1= O .


39

Consequently, cy + 1 is the other solution of our equation. Thus we get the addition and multiplication tables shown in Table 2.1 8. Table 2.18 The Finite Field with 4 Elements

fi N

a

N + l

n+l

0 n+l n+l 0 0

:I;

I

(2

1

ff

0

0+1

0 0 0

ff

n+l

a

a+l N+1 I

(2

We are, in fact, taking the ring of polynomials over 22 and working with them modulo ,r2+ z + I to give the field of order 4. We will write G F [ 4 ]for the field of order 4 (where C:F stands for “Galois field” after the French mathematician Evariste Galois ( 1 8 1 1-1 832)). The integers modulo 71 form a field if and only if 71 is prime. In the same way, if we start from Z p for some prime p , and consider the ring of polynomials Zp[[.c] over Zr),modulo a polynomial f ( x ) ,this forms a field if and only if f ( x ) is irreducible over X p . We will write the entries i n C F [ 4 ]as ordered pairs: (1 I ) = 0 + 1 and (01) = 1 , for instance. The only other thing that we need to know about finite lields is that the multiplicative group of the finite field is cyclic. Thus, if we take all the elements i n the field other than 0, then each element can be expressed as a power of a primitive element of the field. I n the case of C F [ 4 ]the primitive element can be taken as N since o2 = (1 + 1 and o3 = I . For G F [ 5 ]we , can use 2 as the primitive element since 2 2 = 4, 23 = 3 and 24 = 1, or we can use 3 as the primitive element (3’ = 4, 3:’ = 2 and 34 = 1) hut we can not use 4 (since 4‘ = 1). Of course if we represent the entries in GF[4] as ordered pairs then we have the same representation for the lcvels of a four-level factor as we would have from having two factors each with two levels. This can be a useful way of thinking about a four-level factor; we say that the two new factors are thepsedo:fnrctor.y corresponding to the original four-level factor. Three orthogonal contrasts for the four-level factor can be represented by linear. quadratic, and cubic orthogonal polynomials, or they can be represented by the main effects and the two-factor interaction of the pseudo-factors. We will see later that both representations are useful, depending on the circumstances. Armed with these finite fields, we can now give constructions of regular fractions for any symmetric design with a prime or prime-power number of levels. 2.2.4

Fractions for Prime-Power Levels

In this section we will use L to represent either a prime or a prime power. We will write G F [ t ]as 2 is not relevant.

2.3.1

Two Constructions for Symmetric OAs

Various constructions for orthogonal arrays have been found; a nice summary is given in Hedayat et al. (1 999). U7e give two of the most useful constructions in this section. The next construction gives a resolution 3 design with 2t2 treatment combinations for up to 2L 1 factors when I' is odd.

+

w CONSTRUCTION 2.3.1. I f P is on odd prime or prime power then there is an OA[2F2,2P

+ 1;C. 21.

PmoJ Let the elements in C:F[C] = (a0 = O : Q ~ = 1 . 0 2 , . . . . r w e - 1 } . The design is constructed i n two parts. As in Sections 2.2.1 and 2.2.2 we let bl be a vector of length P 2 with the first e entries equal to 0 0 , the next P entries equal to 01 and so on until the final P entries are equal to C P C - I . We let b2 have first entry 0 0 , second entry al and so on until entry e is ar-1 and these entries are repeated in order C times. For the first L 2 treatment combinations we use bl for the levels of the first factor, b2 for the levels of the second factor, and bl aqb2 for the levels of the ( q 2)th factor. q = 1.2, the levels for the remaining l factors, we use b: + aqbl+ bz. q = 0, . . . ,L - 1. For the remaining P2 treatment combinations, we use bl for the levels of the first factor. b2 for the levels of the second factor, bl +crqbz uq; q = 1, . . . , // - 1, for the levels of the next P - 1 factors and Ob: b2, Ob: 8,bl b2 pq, q = 1 , .. . .P - I , for the levels of the final C factors, where we have to determine u1; 4, . . . p l , p2, . . . , pe-', 8. Q 1 82, . . . , Be- 1. We let O be any non-squareelement of GF[P].We let vq = (8- 1 ) ( 4 6 ' 0 ~ ) - ~Oq, = Ooq and pq = ai(8 - 1 ) C 1 ,q = 1:2

+

+

+

+

+ + +

~

~

The result sometimes holds for I even; see Hedayat et al. (1999) (p. 47) for a discussion and another way of constructing OAs with 2n symbols. The smallest design for C a power of 2 that would be given by this construction is an OA with 9 factors each with 4 levels with N = 32. =EXAMPLE 2.3.1.

Let E = 3. Since 2 is not a square in G F [ 3 ]we , let 6' = 2. So we have u1 = (4 x 2 x l ) - ' = 2, v2 = (4 x 2 x 2)-l =: 1,81 = 2 , & = 1, p1 = 4-1 = 1 and p2 = 4-' = 1. Thus the treatment combinations i n the second half are given by bl, b2, bl b2 2, bl 2b2 1, 2b; b2.2b: 2b1+ b2 1 and 2b: bl b2+ 1. The final 18 treatment combinations are given in Table 2.20.

+

+

+

+ +

+ +

+

+

The next result is a special case of a result in Bush ( 1 952) and gives an orthogonal array of strength 4, equivalently a resolution 5 fractional factorial design. CONSTRUCTION 2.3.2.

IfP is o prime or a prinie power nnd P 2 3, then there is on OA(P4.L+ 1,L.4].

44

FACTORIAL DESIGNS

Table 2.20 A Resolution 3 Fractional Factorial Design for Seven 3-Level Factors 0 0 0 1 1 1 2 2 2 0 0 0 1 1 1 2 2 2

0 1 2 0 1 2 0 1 2 0 1 2 0 1 2 0 1 2

0 1 2 1 2 0 2 0 1 2 0 1 0 1 2 1 2 0

0 2 1 1 0 2 2 1 0 1 0 2 2 1 0 0 2 1

0 1 2 1 2 0 1 2 0 0 1 2 2 0 1 2 0 1

0 1 2 2 0 1 0 1 2 1 2 0 2 0 1 1 2 0

0 1 2 0 1 2 2 0 1 1 2 0 1 2 0 2 0 1

Pro08 Use the elements of GF[&] to label the columns of the array and the polynomials of degree at most 3 to label the rows of the array. Suppose that & is the polynomial associated with row i and that field element cy3 is associated with column j . Then the (i, j ) entry is & ( a 3 ) The . final column contains the coefficient of z3 in &. The verification that this array has the desired properties is left as an exercise. 0 4 EXAMPLE 2.3.2. Let e = 3. Then we can use Construction 2.3.2 to construct an OA[81,4,3,4]. There are three polynomials of order 0: 0, 1 and 2. There are 6 polynomials of order 1 : 2 , z 1, z 2, 22, 22 1 and 22 2. There are 18 polynomials of order 2: the 9 with leading coefficient 1 arez2, z2 1, z* 2, z 2+z, z2+ z 1, x 2 + z + 2, x 2 + 2 z , z 2 + 2 2 + 1and z2 22 2. There are a further 9 quadratics with leading coefficient equal to 2. Finally each of these 27 polynomials can have z3 or 2z3 added to it to give the 81 polynomials required in the construction. 0

+ + + +

+

+

+

+

+

2.3.2 Constructing OA[2”; 2”l, 4”z; 41 Addelman (1972) gave some useful constructions for these designs. He described each 4-level factor by three 2-level generators and he gave a set of generators for the 2-level factors. This set had to be such that no linear combination of two or three of the generators in the set was also in the set. We give the six designs that he gave in Table 2.21. EXAMPLE 2.3.3.

To construct an OA[64;23,42;4] we need the generators of length 64. So bl has 32 0s and then 32 1 s, b p has 16 Os, then I6 1 s then 16 0s then 16 1s and so on until bG which alternates 0s and Is. Then we use bg, be and bl b3 b5 b6 to determine the 2-level factors and

+ + +

IRREGULAR FRACTIONS

Table 2.21 Generators for OA[2k: 2 k 1 .4": 41 OA

Two-Level Factors

OA[ 128:29,4:1]

b3. b4. bs, bs, b7. bi b3 b4 b5. bi b3 bs b7. bz b3 bq bs. bz b4 bs b7

(bi. bz, bi

+ b2)

bs, bs, b7.

(bi. b2, bi (b3.b4.b3

+ bz) + bd)

+ + + + + + + +

OA[ 128:2s,,42:4]

+ + + +

bi bi b2

+ b3 + bj + bs. + b4 + be + b7, + b3 + bs + b7

bi bz

+ b3 + b5 + b7, + b4 + bs + b7

OA[128;23,43:4]

b7,

Four-Level Factors

45

46

FACTORIAL DESIGNS

+

use bi, bz, bl b2 and bs, b4, b3 is given Table 2.22.

+ b4 for the two 4-level factors. The resulting design

0

Table 2.22 The OA[64;23,42;4] 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0

0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0

0 0 1 1 0 0 1 1 1 1 0 0 1 1 0 0

0 0 0 0 1 1 1 1 2 2 2 2 3 3 3 3

0 1 2 3 0 1

2 3 0 1 2 3 0 1

2 3

0 1 1 0 0

0 1 1 0 1 0 1 0 0 2 0 1 0 0 3 0 1 1 1 0 0 1 1 1 1

0 0 0 0 0

1 1 1 1 1

0 1

0 0 0 0

1 1 1 1

0 0 0 0 1 1 0 0 1 1

1 1 2 2 2 2 3 3 3 3

2 3 0 1 2 3 0 1 2 3

1 0 1 0 0 I O I O l

1 0 0 0 2 1 0 0 0 3 1 0 1 1 0

I 1 0 1 1 0 1 1 1 1 1 1

0 0 0 0

0 1

2 3

1 0 1 1 1

1 1 0 1 0 1 1 0 1 1

1 1 1 1 1 I 1 1 1 1

1 1 1 1 1 1 1 1 1 1 1 1 I 1 0 1 1 0 1 1 1 1 1 1 1 1 0 I 1 0

0 0 0 0 0 0 0 0 0 0

0 0 0 0 1 1 0 0 1 1

1 1 2 2 2 2 3 3 3 3

2 3 0 1 2 3 0 1 2 3

1 1 2 2 2 2 3 3 3 3

2 3 0 1 2 3 0 1 2 3

2.3.3 Obtaining New Arrays from Old Given an orthogonal array, symmetric or asymmetric, it is possible to get other arrays from it by deleting one, or more, factors and by equating some of the levels for one, or more, of the factors. Sometimes it is possible to replace the entries in one factor by several factors, each with fewer levels, and sometimes it is possible to combine several factors to get one factor with more levels. In this section we will discuss each of these situations in turn. CONSTRUCTION 2.3.3. Collapsing Levels Consider an orthogonal array with a factor with C, levels. Suppose that we want to make this into a factor with 6 , < e, levels. We can do this by changing level C, to 0, e, 1 to 1, C, 2 to 2, and so on, until all the levels in the original factor have been changed to ones for the new factot

+

+

This way of making the changes guarantees that each level of the new factor appears about the same number of times (exactly the same number of times if P, IC,) and the properties of the original array guarantee that the final array is an OMEP. But any collapsing that results in levels appearing about the same number of times is just as good. This procedure is almost impossible to reverse. EXAMPLE 2.3.4.

Consider the array in Table 2.20. Suppose that we want an array with 7 factors, 3 with two levels and 4 with three levels. Then we could collapse the levels in the first three factors to get the OMEP shown in Table 2.23. Note that the first three factors have level 0 repeated 12 times and level 1 repeated 6 times. 0 The next construction only applies to orthogonal arrays of strength 2.

IRREGULAR FRACTIONS

Table 2.23

47

A Resolution 3 Fractional Factorial Design for Three 2-Level Factors and Four

3-Level Factors 0 0 0 I I 1 0 0 0 0 0 0 1

0 1 0 0 1 0 0 1 0 0 1 0 0

1 1

1 0 0 0

0 0 1 0

0 1 0 1 0 0 0 0 1 0 0 1 0 1 0 1 0 0

0 2 1 1 0 2 2 1 0 1 0 2 2 1 0 0 2 1

0 1 2 1 2 0 1 2 0 0 1 2 2 0 1

2 0 1

0 1 2 2 0 1 0 1 2 1 2 0 2 0 I 1 2 0

0 1 2 0 1 2 2 0 1 1 2 0 1 2 0 2 0 1

W CONSTRUCTION 2.3.4. Expansive Replacement Suppose that we have an orthogonal array of strength 2 in which there is one,factor u i / h levels. Sicppose that there is an orthogonal array of strength 2 which has t1 runs. Label these rirn.s,frorn0 to P1 -- 1. Then we ran replace each level of the factor with PI levels in the first array by the corresponding run of the second array. 01

The name arises because the new array has more factors than the original array. WEXAMPLE2.3.5.

Consider the array i n Table 2.24. Suppose that we want to replace one of the factors with the runs from the OA[4,3,2,2]. The nins of this array are given in Table 2.25. The resulting OA[ 16;2,2.2,4,4,4,4;21 is given in Table 2.26. If we make this replacement for all five of 0 the 4-level factors, then we get the array in Table 2.27. The most common examples of expansive replacement are replacement of a factor with 4 levels by three factors each with 2 levels and replacement of a factor with 8 levels by seven factors each with 2 levels. It is sometimes possible to replace several factors with one factor with more levels. Before we describe this contractive replacement we need to define the idea of a a tight or saturated orthogonal array. In a tight or saturated orthogonal array with k factors, N = 1 C,"=, ( k q - 1) Thus we see that there is no room for any more factors.

+

CONSTRUCTION 2.3.5. Contractive Replacement Let A be nn O A [ N ;t 1 ( 2 : . . , p k ; 21 such that thejrst s colirrnns of A,forrn N/Nl copies of an OA[n', : PI E 2 , . . . , E,; 21, B say, that is tight. Use 0 , 1, . . . . ilil - 1 to label the rows of B, and theri use these labels to replace tlze$rst s. colurnns of A . The resulting design is 0 an OA[:V; N I .t q +~. l . . F k : 21. ~

~

.

48

FACTORIAL DESIGNS

Table 2.24 An OA[16,5.4,2] 0 0 0 0 1 1 1 1 2 2 2 2 3 3 3 3

0 1 2 3 0 1 2 3 0 1 2 3 0 1 2 3

0 1 2 3 1 0 3 2 2 3 0 1 3 2 1 0

0 2 3 1 1 3 2 0 2 0 1 3 3 1 0 2

0 3 1 2 1 2 0 3 2 1 3 0 3 0 2 1

Table 2.25 An OA[4,3,2.2] 0 0 1 1

0 1 0 1

0 1 1 0

Table 2.245 An OA[ 16;2.2,2.4,4,4,4:2] 0 0 0 0 0 0 0 0 I 1 1 1 I 1 1 1

0 0 0 0 1 1 1 1 0 0 0 0 1 1 1 1

0 0 0 0 1 1 1 1 1 1 1 1 0 0 0 0

0 I 2 3 0 1 2 3 0 1 2 3 0 1 2 3

0 1 2 3 1 0 3 2 2 3 0 1 3 2 1 0

0 2 3 1 1 3 2 0 2 0 1 3 3 1 0 2

0 3 1 2 1 2 0 3 2 1 3 0 3 0 2 1

IRREGULAR FRACTIONS

Table 2.27 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1

0 0 0 0 1 1 1 1 0 0 0 0 1 1 1 1

0 0 0 0 1 1 1 1 1 1 1 1 0 0 0 0

0 0 1 1 0 0 1 1 0 0 1 1 0 0 1 1

0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1

0 1 1 0 0 1 1 0 0 1 1 0 0 1 1 0

49

An OA[I6.15,2,2] 0 0 1 1 0 0 1 1 1 1 0 0 1 1 0 0

0 1 0 1 1 0 1 0 0 1 0 1 1 0 1 0

0 1 1 0 1 0 0 1 1 0 0 1 0 1 1 0

0 1 1 0 0 1 1 0 1 0 0 1 1 0 0 1

0 0 1 1 1 1 0 0 0 0 1 1 1 1 0 0

0 1 0 1 1 0 1 0 1 0 1 0 0 1 0 1

0 1 0 1 0 1 0 1 1 0 1 0 1 0 1 0

0 1 1 0 1 0 0 1 0 1 1 0 1 0 0 1

0 0 1 1 1 1 0 0 1 1 0 0 0 0 1 1

uEXAMPLE2.3.6.

Consider the design in Table 2.26. The first three columns form four copies of a tight OA[4.3,2.2]and so replacing 000 with 0,011 with I , 101 with 2 and I 10 with 3 gives the 0 OA[ I6,5,4,2] in Table 2.24. The same construction will work if the array A could have columns adjoined so that

B would be tight. The construction does not work if the original array is not tight, or cannot be made tight, or if only some columns o f the tight subarray are replaced, as the next example shows. uEXAMPLE2.3.7.

Consider the first two columns of the orthogonal array in Table 2.26. Then these two columns form 16/4=4 copies of an OA[4,2,2,2], although it is not tight since 1 t (PI - 1)

+ ( P 2 - 1) = 1 + 1 + 1 = 3 # 4.

If we replace 00 by 0, 0 1 by I , 10 by 2 and 1 1 by 3 we see that the 0s in the third column of the original array only appear with 0 and 2 in the proposed new 4-level column. This is because the third column is the binary sum of the first two columns and so its levels are not independently determined. The first three columns of the array do form a tight OA [4,3,2,2]. By juxtaposing several copies of an array it is possible to get a larger array with k iactors.

+1

CONSTRUCTION 2.3.6. Adding One More Factor Take any OA"; P1 f z ; . . . , Pk3 t ] and write down copies of the array, one above the other; and then adjoin a ( k + 1 ).st,factor with N copies of 0, then N copies of I , and so on, iiiitiltherearen:copiesqfFk+1 - I . T h e r e s i i l t i s a n O A [ N xP k + l ; [ l . i ! 2 , . . . . c k , i ! k + j : f ] .

50

FACTORIALDESIGNS

The levels of the last factor appear equally often with every pair of levels for any other two factors and so the array has strength 3 with respect to the last factor; this property is useful when constructing an OA to estimate two-factor interactions all of which involve one factor.

n EXAMPLE 2.3.8. Consider the OA[4,3,2,2] given in Table 2.25. By writing down three copies of this array and adjoining one further column for the three-level factor, we obtain the OA[ 12;2,2,2,3;2] in Table 2.28. 0 Table 2.28 An OA[ 12;2,2,2,3:2] 0 0 1 1 0 0 1 1 0 0 1 1

0 1 0 1 0 1 0 1 0 1 0 1

0 1 1 0 0 1 1 0 0 1 1 0

0 0 0 0 1 1 1 1 2 2 2 2

The next construction extends this idea by adding several factors simultaneously. H CONSTRUCTION 2.3.7. Juxtaposing Two OAs Takean OA[N1;e l l , P12, . . . , e l k ; t l a n d a n OA[Nz;P21, P 2 2 , . . . , .&:tlandwritedown Nz copies of thejirst array, one above the othel: Adjoin thejrst row of the second array to the Jirst ropy of thejirst array, adjoin the serond row of the serond array to the second copy of the Plz,. . . , t l k , E z 1 , & ~ ,... , C z k ; t ] . 0 jirstarrayandsoon. TheresultisanOAININz;ell,

In an OA obtained from Construction 2.3.7 there are several sets of three factors in which every triple of levels appears equally often. These include every set of two factors from the original OA and one factor from the adjoined OA and every set with one factor from the original OA and two factors from the ad,joined OA. uEXAMPLE2.3.9.

Consider the OA[4;3,2;2] given in Table 2.25. Use this as both the original OA and as the adjoined OA. Then we get the OA[ 16;6,2;2] in Table 2.29. Notice that the only two sets of three factors which do not contain all the triples of levels are factors 1, 2, and 3 and factors 4, 5, and 6. 0

n CONSTRUCTION 2.3.8.

Obtaining One Factor with Many Levels

Take any orthogonal array O A [ N ;e l , Cz, . . . ,C k ; t ] and write down C copies of the array, one above the other, and then leave the levels of the kthfactor unaltered in thefirst copy ofthe array, use C k different levels in the second copy of the array, use afrcrtherPk levels in the third ropy of the array and so on. The resulting array is an OAINP; f 1, pz . . . , Pk-1. P&: t]. 0

IRREGULAR FRACTIONS

51

Table 2.29 An OA[16;6.2;2] 0 0 0 0 0 0 0 1 1 0 0 0 1 1 0 0 1 1 0 0 1 1

0 1 0 1 0 1 0 1 0 1

1 0 0 1 1 0 0 1 1 0

0 0 0 0 0 0 1 1 1 1

0 0 0 1 0 1 1 1 1 0 1 1 1 1 0 1

0 0 1 1 1 1 0 0 0 0 1 1 1 1

0 0 I 1 1 1 1 1 1 1 0 0 0 0

EXAMPLE^.^.^^.

The OA[ 12;2,2,6;2] in Table 2.30 is obtained by writing down three copies of the OA[4,3,2,2] given in Table 2.25 and replacing the final column with three distinct sets of two symbols for the final factor. [7 Table 2.30 An OA[ 12;2,2.6:2] 0

0

0 1 1 0 1 1 0 0 0 1 .

0 1 1 0 2 3

1

0

3

1

1

2

0 0

0 4 1 5 1 0 . 5 1 1 4

The next construction shows how to remove an unrealistic (or unwanted) treatment combination from a fractional factorial. CONSTRUCTION 2.3.9. Avoiding an Unrealistic Treatment Combination Given an orthogonal array OAlR:P I ,l z , . . . , & I , e k ; t ] which contains an unreali.stic treatment combination, the treatment combination can be removed by adding another treatment conibination to even. treatnient coinbination in the arm): This addition is done component-wise mod f 4 in position q.

52

FACTORIAL DESIGNS

The treatment combination to add needs to be chosen so all the resulting treatment combinations are acceptable; often such a treatment combination can only be found by trial and error. EXAMPLE 2.3.11. Suppose that we want an OA[12;2,2,6;2] without the treatment combination OOO. Then we need to add a treatment combination to those in the array in Table 2.30 so that 000 is avoided. This means that we need to add some treatment combination that is not the negative of any of the treatment combinations that are already there. So we try using 1 13. This gives the array in Table 2.31. 0

Table 2.31 An OA[ 12;2,2,6;2] without 000

2.3.4

1 1 0

1 0 1

3 4 4

0

0

3

1 1 0 0 1 1 0 0

1 0 1 0 1 0 1 0

s 0 0 s 1 2 2 1

Exercises

1. Use Construction 2.3.2 to construct an OA[8 1,4,3,4].

2. Prove that the array constructed in Construction 2.3.2 has the desired properties. 3 . Consider the design in Table 2.32. (a) Verify that the design is of resolution 3 . (b) Is it resolution 4? (c) Show that the final three columns can be replaced by one column with 4 levels. (d) Are there any other sets of three columns for which you could make this replacement? (e) In particular, can you write the design as a resolution 3 array with two factors each with 4 levels? 4. Construct an OA[32:24,4;41.

OTHER USEFUL DESIGNS

53

Table 2.32 An OA[8,6.2.2] 0 0 0 0 0 0 0 1 1 0 1 1 1 1 0 0 1 1

2.4

0 1 0 1 0 1

1 0 1 0 0 1

1 1 1 1 0 0

0 1 1 0 1 0

1 0 0 1 1 0

OTHER USEFUL DESIGNS

There are various other combinatorial designs that have been used in the construction of stated preference choice experiments. We collect the definitions here for convenience. Consider a set of 2' items. From this set of 7' items construct b subsets, or blocks, each with z( distinct items. Suppose that any two items appear together in exactly X of the h blocks. Then the set of 6 blocks form a balanced inromplete block design (BIRD). Let r, be the number of times that item 7; appears in the BIBD. If we count the pairs that item '1 appears in the BIBD we have that 7; appears in r , blocks and there are i l ~- 1 other items in the blocks. So there are r i ( u - 1) pairs involving item ?;. On the other hand there are - 1 other items in the RIBD and item 7; appears with each of these X times in the RIBD. So there are X ( r - 1) pairs involving item Yi. Equating we get that r t ( u - 1) = X(73 - 1 ) and so we see that all of the items appear equally often in the BIBD. We call this replication number r and talk about a (v. h: r. u;A) BIBD. If 71 = b then r = u and the design is said to be a symmetric BIBD (SBIBD), written (v, r, A). Tables of balanced incomplete block designs may be found in Mathon and Rosa (2006) and Abel and Greig (2006). EXAMPLE 2.4.1. Let 1' = 7 and consider the blocks in Table 2.33. These blocks form a BIBD on 7 items with 7 blocks each of size 3 and with each item appearing in 3 blocks. There is a unique block which contains every pair of items so A = 1.

Table 2.33 The Blocks a (7,3,1) BIBD 1

2

3

1 1 2 2

4 6 4 5

5 7 6 7

3 3

4 5

7 6

One easy way to construct BlBDs is to use diflerence sets Let X = ( ~ 1 . 2 2 , . . n u } be a set of integers modulo v . Suppose that every non-zero value mod 2' can be represented

54

FACTORIAL DESIGNS

as a difference x, - x3, x,,x3 E X, in exactly X ways. Then X is said to be a difference set.

(I).7 ~ A) ,

W EXAMPLE2.4.2. Let v = 7 and = 3 and let X = {0,1,3). Then X is a (7,3,1)-difference set since each of the non-zero integers mod 7 can be represented as a difference in exactly one way using the elements from X . In fact we have 1 = 1 - 0, 2 = 3 - 1, 3 = 3 - 0, 4 = 0 - 3, 5 = 1 - 3 and 6 = 0 - 1. 0 We give some small difference sets in Table 2.34. More extensive tables of difference sets are given by Jungnickel et al. (2006). Table 2.34 Some Small Difference Sets

v u,x

Set

7 3 7 4 13 4 21 5 II 5 11 6 IS I

O.I.2.4 0,1,3,9 3,6.7,12, I4 1.3.4s.9 0,2,6,7,8,10 0,1,2,4,5,8,10

1 2

I

1

2 3 3

Of course, if X = { X I ,xz,. . . . 2,) (21

0,1,3

is a (v.u,A) difference set then so is

+ 1,zz + 1 , .. . + 1). ,xt'

+

It is also true that X = (0.1, . . . , v - 1)\X is an (v,v - 7 ~ v, - 271 A) difference set. We can see that this is true by counting the number of times that each non-zero value mod u appears as a difference. The total number of differences in X is (21 - .)(ti - u - 1). The total number of differences in X is u(u- I ) and each of the 71 - 1 non-zero values mod v appear as a difference X times so u ( u - 1) = X ( w - 1). Thus we see that (2' - u)(u -

u - 1) = v(v

- u(v - 1 - u ) 1) - 2W - U2' - +1L(P/ + 1) = V ( V - 1) X(7l - 1) 221 - 21L7) = V(7l - 1) x ( 7 1 - 1) - 2 7 / ( 7 1 - 1) = (2' - l)(v + x - 2u).

=

Zl(7'

-

1 - 7L)

-

+ +

+

Since {0,1,.. . , V - 1) is a ( v ,V , v) difference set the result follows. It is not essential for a difference set to be defined over the integers mod 11. Any group can be used although the only extension that we will make here is to give one difference set defined on ordered pairs where differences on each element of the pair are evaluated mod 4. Thus there are 16 levels in total and each is represented by an ordered pair. If u = 6 then theorderedpairs { ( O , O ) , ( O , l ) , ( l , O ) , (1,2), (2,0), ( 2 , 3 ) ) forma(l6,6,2)difference set. For example we see that the element (1,3) arises as a difference twice: ( 2 , 3 ) -- (1,O) and ( 1 , O ) - ( 0 , l ) . A further extension is to have several sets and look at the differences across all the sets. Formally suppose that X,= { z , ~ ~, ~ 2. . ,, x. t U }is a set of integers modulo v. Then

TABLES OF FRACTIONAL FACTORIAL DESIGNS AND ORTHOGONAL ARRAYS

55

X I . X z . . . . , X f form a (v. u,A) difference fanlily if every non-zero value mod 1 ’ can be represented as a difference x t n - x,b, x,, , T C , ~E 1 5 i 5 ,f, in exactly A ways.

x,,

BEXAMPLE2.4.3.

The sets (0,1,4) and (0,2,7) form a (1 3,3,1) difference family.

0

Table 2.35 contains some small difference families. Table 2.35 Some Small Difference Families

v

11

x

13 3 1 19 3 1 I6 3 2 11 4 6 IS 4 6

19 4 2 13 s 5 17 5 S

Sets 0,1,4 0.1.4 0.1.2 0,1,8,9 0.1.2.3 0.3.6.9 0.1.3,12 0.1.2,4,8 0.1,4.13.16

0,2.7 029 02.8 0.2S.7 0,2,4,6

0,s.1 1 0,3,7

0.4.7 0,S.lO 0,2,3.S 0.43.9 0,4,8.120.1,8.9 0.1.5.10 0.2,S.IO 0.1,5,13 0,4,6,9 0,1,3,6.12 0,2,5,6.10 0.3,5,12,14 0,2,8,9.lS 0.6,7,10,11

0,1,4,S

A Hadamard matrix of order h is an h x h matrix with entries 1 and -1 satisfying

HH’= It,. A table of Hadainard matrices is maintained by Sloane (2006a). mEXAMPLE2.4.4.

Here is a Hadamard matrix of order 4.

[ i -; -; I;] 1

1

-1

2.5

1

-1

1

0

TABLES OF FRACTIONAL FACTORIAL DESIGNS AND ORTHOGONAL ARRAYS

There are two extensive tables of orthogonal arrays available on the web. One is the table of orthogonal arrays maintained by Sloane (2006b). This website uses “oa.N.k.s.t.name” to denote an orthogonal array with A; runs, k factors, s levels (for each of the factors), and strength t (equivalently resolution t + I). This site sometimes gives mathematically inequivalent designs with the same parameters and it makes no claim that it lists all possible parameters, even for small designs. It has many designs of resolution 3 and some designs for larger resolutions. The other tables are maintained by Kuhfeld (2006) and list all parent designs of resolution 3 with up to 143 treatment combinations as well as many designs with more. A parent design is an orthogonal array i n which no further contractive replacement is possible. As well as listing the parent arrays, Kuhfeld has a list of the number of parent designs for a given value of N and the number of designs that exist for that value of ,V.For example, there are 4 parent designs and 4 designs with N = 12 (OA[ 12.1 1,2,2],

56

FACTORIALDESIGNS

OA[ 12;2,2,2,2,3;2], OA[ 12;2,2,6;2] and OA[ 12;3,4;2]), while there are 2 parent designs with N = 16 (OA[ 16;2,2,2,2,2,2,2,2,8;2] and OA[16,5,4,2]) but 7 designs obtainable from them by expansive replacement. This site has no designs of resolution other than 3. When trying to find a design that you need from these or any other tables, remember that you can omit factors without affecting the resolution of the design. Frequently, designs that are not tight are not listed and such designs are often what is required. Thus you need to think in terms of collapsing levels and omitting factors from the tabulated designs to get the ones that you want.

Exercises

2.5.1

I . Go to the website http://www.research.att.com/-njas/oadir/and find a design with k = 5 factors, four with 2 levels and one with 3 levels. How many treatment combinations does it have? How would you define the main effects for each of these factors? 2. Consider the column (1, 1, 1,0,1,0,0)’. Obtain 6 further columns by rotating one

position for each new column. Adjoin a row of 0’s. Verify that the resulting array has 7 binary factors and is of resolution 3. 3. Use one of the websites mentioned above to find an OA of strength 2 for eight 4-level factors and an 8-level factor in 32 runs. Now make an OA with 288 runs with one

more factor with nine levels. Show how to use this design to get an OA with two 2-level factors, four 4-level factors, an 8-level factor and one factor with 36 levels still in 288 runs. 2.6

REFERENCES AND COMMENTS

Although used in some agricultural experiments from the mid-1800s. the first formal exposition of factorial experiments was given by Yates (1935) and they appeared in Fisher (1 935).

There are various books that develop factorial designs for use in comparative experiments including Kuehl (1999), Mason et al. (2003) and Montgomery (2001). These books also discuss the derivation of orthogonal polynomials, for equally spaced levels, whose use in model fitting was first pointed out by Fisher. The links between fractional factorial designs and finite geometries were pointed out by Bose and Kishen (1940) and exploited by Bose ( I 947). Much of the early work in the area is summarized in Raghavarao (1 971) and Raktoe et al. (1981). There is a close relationship between fractional factorial designs and orthogonal arrays, and an extensive treatment of results pertaining to orthogonal arrays can be found in Hedayat et al. (1999). They give a number of constructions that rely on the structural properties of the original OA to allow further factors to be added. The simplest of these ideas was presented in Section 2.3.3. The initial theoretical development of regular fractional factorial designs appeared in Fisher (1 945) and Finney ( 1 945). Dey ( I 985) provides a number of constructions for fractional factorial designs and OMEPs. There has been some debate about the appropriate contrasts to use in a fractional design and some of the issues are addressed in Beder (2004) and John and Dean ( 1 975).

CHAPTER 3

THE MNL MODEL AND COMPARING DESIGNS

As we explained in Chapter 1, a stated preference choice experiment consists of a finite

set of choice sets and each choice set consists of a finite number of options. The options within each choice set must be distinct and they must be exhaustive, either because there is an “other” option, a “none of these” option or because the subjects are asked to assume that they have narrowed down the possible options to those given in the choice set. This final assumption is often described by saying that the subjects have participated in a,forced choice stated preference experiment. We assume that each subject chooses, from each choice set, the option that is “best“ for them. The researcher knows which options have been compared in each choice set and which option has been selected but has no idea how the subject has decided the relative value of each option. However, the researcher assumes that these relative values are a function of the levels of the attributes of the options under consideration, some of which have been deliberately varied by the researcher. In this chapter we begin by defining utility and showing how a choice process that bases choices on the principle of random utility maximization can result in the multinomial logit (MNL) model for a specific assumption about the errors. We then derive the choice probabilities for the MNL model. Next we consider the Bradley-Terry model which arises when all choice sets have only two options, before looking at the MNL model for any choice set size. We assume that the MNL model is the discrete choice model that is the decision rule sub.jects are employing. The results on optimality that we present will depend on this assumption. Should a different model be used, then different designs may well prove to T/ir Cori,,r,srrri~~rion ofOp/irnn/ Srnred Choice b,kperirneril,s. By D.J . Street and L. Burgesh Copyright @ 2007 John Wiley & Sonr. Inc.

57

58


be optimal. We also discuss how to decide objectively which of a set of proposed choice experiments is the best one to use, from a statistical perspective, for a given situation. At times in this chapter, and indeed throughout the rest of the book, we will talk about options that are described by the levels of k attributes. These attributes are precisely the same as the factors of the previous chapter.

3.1 UTILITY AND CHOICE PROBABILITIES In this section we define utility and show how a choice process that bases choices on the principle of random utility maximization can result in the MNL model for a specific assumption about the errors.

3.1.1

Utility

Train (2003) has defined utiliry as “the net benefit derived from taking some action”; in a choice experiment, we assume that each subject chooses the option that has maximum utility from the ones available in each choice set. Thus each subject assigns some utility to each option in a choice set and then the subject chooses the option with the maximum utility. If we let U,, be the utility assigned by subject a to option J , J = 1 , . . . , m, in a choice set with m options, then option i is chosen if and only if U,, > U,, V J # i. The researcher does not see the utilities but only the options offered and the choice made (from each of the choice sets). These options are usually described by levels of several attributes. The systematic component of the utility that the researcher captures will be denoted by V,,, and we assume that U,, = V,, el,, where cJa includes all the things that affect the utility that have not been included in V,, . Thus the cJa are random terms, and we get different choice models depending on the assumptions that we make about the distribution of the

+

E p .

We know that Pr(option i is chosen by subject a ) = PT(U,, > U,, V J # i) = PT(v,, €z, > V,, €,a

+

=

Pr(K,

-

V,, > f,,

+

v j # 2)

- €%,

vj

# i).

If the E ~ are , independently identically distributed extreme value type 1 random variables then the resulting model is the multinomial logit (MNL) model. This assumption on the e J a is equivalent to assuming that the unobserved attributes have the same variance for all options in each choice set and that these attributes are uncorrelated over all the options in each choice set (Train (2003)). Train (2003) gives an example for choosing travel options where this independence assumption is not reasonable: If a subject does not want to travel by bus because of the presence of strangers on the bus, then that subject is also unlikely to choose to travel by train; and discusses other models that have been proposed which avoid this independence assumption. These models are the generalized extreme value (GEV) models, the probit model and the mixed logit (ML) model. In the generalized extreme value models, the unobserved portions of the utility are assumed to be jointly distributed as generalized extreme value. Consequently, correlation in the unobserved attributes is allowed, but this model collapses to an extreme value model if the correlation is 0 for all attributes. In the probit model, the

59

UTILITY AND CHOICE PROBABILITIES

unobserved attributes are assumed to be normally distributed. This means that unobserved attributes can be assumed to be jointly normal and be modeled with any appropriate correlation structure over both options and time (for example, with panel data where the same individuals are asked to respond on several successive occasions). Of course it is rarely appropriate for a price attribute to come from a normal distribution since few people want to pay more for the same or worse features (but consider those who like to emphasize their wealth through corispicuous consumption). This imposes some limitations on the application of the probit model. In the mixed logit model, the unobserved attributes can be decomposed into two parts: one with the correlation, which may have any distribution, and one which is extreme value. Variations on these models, such as nested logit, are also possible. While there are many different choice models, we will only discuss designing choice experiments for the multinomial logit model in this book. 3.1.2

Choice Probabilities

We assume that each unobserved component of the utility F,, is independently identically distributed Type 1 extreme value, sometimes called the Gumhel distribution. It has density f ( E ) = p-te-e

I

,

--35

< F < 92

and cumulative distribution function

F(E) = e - e - ‘ . We now derive the choice probabilities using the argument given in Train (2003, pp. 4044,78-79). From the results above, we have that

P,, = Pr(option i is chosen by subject a ) = Pr(&

-

l/Ja + E,, >

b’l

#

i)

If we regard E,, as given, then P,, is just F(t,, + V,, - 4,) = P - ~ - ( “ ’-‘’ - ‘ I ” ) , again for all J # z. Now the F ’ S are independent and hence the cumulative distribution over all J # z is lust the product of the cumulative distributions over all J # 2 . This gives

Pr(optioi1 i is chosen by subject ajt,,) =

n

-(,,,+L,,

c - ~

‘,,)

3#,

In practice, we do not actually know the values by their density, to give

F,,

and so we need to integrate out

weighting

60


+

To evaluate this integral, observe that V,, - V,, = 0;so E , , V,, - V,, = fz,. can remove the restriction on the product by using the last term. This gives

I/,m

J c,,, = -m

3

1

- V?,Y)

-€

10

Hence we

dE,, .

Now we let z = e-"~ 1 2 , 0 0i ,~ i P . 0 1 ) .

The treatment combinations, and the corresponding xi and yz,are given in Table 3.5. Table 3.4 Lnconstrained Representation of Treatment Combinations Treatment Combination 0.0 0.1 02 1 ,o 1.1

I .2

Xi

I ,o,1 ,o,o,1,o,o,o,o,o 910 I .0.0,1,0,0,1,0,0,0,0 010 I.0.0,0,1,0,0,1,0,0,0 01 0 0,I,1,0,0.0,0.0,1,0,0 3 1 I 0,I.0,1,0,0,0,0,0,1,0 41 1

0,1.0,0.I.0,0,0,0,0,1 91 1

We will be using the first representation in this book.

+ 0+ + + + + +42 o + + + + + 02

1312 00

02 1

3 1 2 01

$2 2

8 1 2 02 3 1 2 10

82 1

312 11

02 2

1312 1 2

0

72


Table 3.5 Constrained Representation of Treatment Combinations Treatment Combination

YP

Xa

Questions about the main effects of attribute q are answered by considering contrasts of the bq,f,,questions about the two-factor interaction of attributes q1 and qz are answered by considering contrasts of the /3qlq2,fql f,,z and so on. If we want to develop the best design to test if some particular effects are zero, we do this by specifying which effects we want to test are zero, and which effects (if any) we are going to assume are zero, and then we find the parameter estimates and determine the asymptotic properties of these estimates. We write the p contrasts that we want to test as the rows of the matrix Bh. We choose the coefficients in the contrast matrix Bh so that BhBk = I p where Z p is the identity matrix of order p . That is, we say that the matrix Bh is orthonormal. If any row of Bh is not already of unit length then we divide the entries in that row by the square root of the sum of the squares of the entries in that row. All contrasts that we assume are 0 appear as rows of the orthonormal matrix B, which we will assume has a rows. Finally, B, contains any remaining contrasts from the complete set of L - 1 independent contrasts. The next example illustrates these ideas for three binary attributes. mEXAMPLE3.2.6. Let k = 3 and let = e 2 = e3 = 2. So the eight items are represented by 000,001,010, 01 I , 100, 101, 1 10 and 1 I 1, in lexicographic order. Then Toil =

Pi,o

+ D2.1 + P 3 , i + P 1 2 , o i + Pi3.01+ Pz3.11 + P123,oii

The constraints that we impose are Pq,o

bq1qz,OO+Pqlqz,Ol

~123,000fb123.001

+ Pq,l

= 0, 9 = 1 , 2 , 3 .

= P q 1 q ~ , 1 0 f P q ~ q ~ ,=1O1,

= ~ 1 2 3 , 0 1 0 + 4 1 2 3 . 0 1 1=

= 1,2;1,3:2,1;2,3;3,1:3,2,

fl123,100+/j123,101

=

~123,110+~123,111 =

0,

and similarly for each of the other two attributes for the components of the three-factor interaction. Suppose that we want to test if the main effects are 0. Then

Bh = Observe that RhBk = 13.

-1 -1 -1

-1 -1 -1 1 1 -1 1 1 -1 -1 1 -1 1 -1 1

1 1 1 1 -1 1

THE BRADLEY-TERRY MODEL

73

Suppose that we do this assuming that the three-factor interaction is 0. Then 1

B ----1 -

1 1 -1

2J;T

1 -1

-1

11

That means that we have made no assumptions about the two-factor interaction contrasts;

[

SO

B

-L 2J;T

1

1 -1 -1 1 -1 1 -1 1 -1 -1 1

-1 -1

-1 1 -1

1 -1

-1

1 . 1 l 1

The likelihood function itself is independent of any constraints that we impose on the parameters but the solution will have to reflect the constraints. We summarize the results in Bradley and El-Helbawy (1976) below. We want to maximize the likelihood equation

subject to the normalizing constraint C ,ln(nt) = C ,yl = 0 and the assumed constraints Boy = 0, where 0, is a column of 0s of length a. Working with the log-likelihood and using Lagrange multipliers we maximbe

subject to L y1 = 0 and R a y = 0,. As before we assume that every item has been chosen at least once. To maximize the function we differentiate with respect to each T?in turn. For T~this gives

= z 1 , z = ( ~ 1 . ~ 2 .,. ., ZL)' and K = Let us7 - C,+ the derivative in matrix notation, we have

z

( ~ 2 . ... , K , + I ) ' .

Then, writing

+ K 1 j L + BAK = Or.

Pre-multiplying by j ; we get

jiz Since j i z have

=

+ K j j i , j 1 2 + j',BAK

K

=

01,.

sly - s N and j;,Ba = 0, we have that L K = ~ 0 and 5ince R,RA = I, we

B,z Hence

=

-I?,

+ K = OL.

z and z

-

B;&z = ( I - B : B a ) z = 01,.

As before these equations are solved iteratively. Their convergence is guaranteed, provided that there are no items that are never chosen; see El-Helbawy and Bradley (1977) for details.

74


What about the distribution of these constrained maximum likelihood parameter estimates? To find this we use the ideas of El-Helbawy and Bradley (1 978) who apply the results of Bradley and Gart ( 1 962) to the parameters in information matrix C of

0

[ E: ]

7 = BhTy = e. m e n the

6 has entries Cuqgiven by

Thus we have

C = BhTA(T)BkT. Unless we need to emphasize the value of n we are considering, we will write A for A ( n ) from now on. We are really interested in BhT of course, and these are the first p entries of 6. The variance-covariance matrix of

0 6=

[ i:; ]

is c-1. If we write

then the variance-covariance matrix of m B h T is given by ;3 ;( provided, of course, Chr = O p , L - l - p - a . If ChT # O p , ~ - l - p - a then the variance-covariance matrix of m B h T is obtained from the principal minor of order p of C - * ; that is, the submatrix of order p starting in position (1 , l ) . EXAMPLE3.2.7.

Consider the situation of Example 3.2.6. Suppose that we use the pairs (000, 1 1 I ) , (001, 1 lO),(OlO, I O l ) , (01 I , loo), (000,Ol I), (000,101),(000,1 lo), (1 1 I , 001),(11 l , O l O ) , and (1 1 1 , 100) and calculate A assuming that n = j,. We get

A=-1

40

4 0 0 -1

0 2 0 0 0 0 -

0 -1 -1-1 -1 -1

0 -1 0 -1 -1 -1 0 0 0 0 - 1 - 1 2 0 0 - 1 0 - 1 0 2 - 1 0 0 0 0 - 1 2 0 0-'1 1 0 0 2 0 0 0 0 0 0 2 0 -1 0 -1 0 0 4

and 4 1 1 4 1 1 . . . 0 0 0 0 0 0

1 : 1 : 4 : . . . 0 : 0 : 0 :

o o o o o o o o o

. . . . 2 1 1 1 2 1 1 1 2

. -


75

AC-', where

The variance-covariance matrix for m l ? h l . ? is 100 -20 -20 : -20 100 -20 : -20 -20 100 :

0 0 0

0 0 0

0 0 0

. . . . . . . . . . . . . . . . . . 0 0 0

0 0 0

0 0 0

270 -90 -90

: : :

-90 270 -90

-90 -90 270

Thus we see that the elements of Bh? and l??? are independently estimated (since the covariance of any element of Bh? and B,.? is 0) and so the variance-covariance matrix of m B I L ? is the principal minor of order 3 of C-l. 0 However it is not always the case that the elements of estimated, as the next example shows.

l?h?

and B,.? are independently

mEXAMPLE3.2.8.

Let k = el

= E2 =

2. Suppose that we want to test hypotheses about the main effects so

T1

Rh = -

2

-1

- l1 -1

l l 1

and we do not want to make any assumptions about the two-factor interaction effect. Then 1

l? - - [ l

-1

7'-2

-1

11.

Suppose that we use the pairs (00, I I ) , (01, lo), and (00,Ol). Then 2

A=-

-1

0

-1

1

If we assume instead that

B,+ = 0. then

Hence we see that with these choice sets the variance estimate of the estimate of the main effect of the second factor, (711 + ;i~l - (710+ 7 0 0 ) ) / 2 ,depends on the assumption, if any, made about the two-factor interaction. 0 The off-diagonal elements of the variance-covariancematrix C-' give the covariance of pairs of contrasts in m B h ? . If one of these entries is 0 then the corresponding contrasts are independently estimated. So if the variance-covariance matrix is diagonal, then all

76


pairs of contrasts are independently estimated. Of course the inverse of a diagonal matrix is diagonal; so we prefer designs in which the C matrix is diagonal. If the attributes that are used to describe the options in a choice experiment have more than two levels then there are often several different ways in which the contrasts in Bh can be represented. We can move between these different representations of the contrasts by multiplying appropriate matrices together. Thus we are not concerned if the C matrix has non-zero off-diagonal entries that correspond to contrasts for the same effect, main or interaction. If we think of C as being made up of block matrices of order (& - 1) x (& - l ) ,for - 1) x (lqJ - 1) (Pq4 - 1)for the twothe main effects, block matrices of order (tql- 1)(lq2 factor interaction effects, q l , 92, q3,q4 = 1 , 2 , . . . , k, and so on for higher-order interaction effects, then we want the off-diagonal blocks to be a zero matrix of the appropriate size, but the diagonal blocks are not restricted in their form. In the case when all the off-diagonal block matrices are zero, we say that C is block diagonal and we note that since the inverse of a block diagonal matrix is block diagonal, the covariance of any two contrasts from different attributes is 0. If we want the components of the contrasts for a particular effect to be independent, then we would have to choose a suitable set of contrasts to achieve that. While this may be possible, it depends on the choice sets used in the experiment and it is sometimes true that the contrasts that result do not have an easily interpretable meaning. Can we say under what circumstances the estimates of the variance-covariance matrix of BhT is independent of the form of B,? The short answer is yes, when Chr = 0, and we discuss this further in Chapters 4,5 , and 6. It is not essential that the elements in Bh be independent contrasts, or even that they be contrasts. For example, if we want to work with the second representation of P in a main effects only setting, then we merely need to use as Bh a matrix which calculates Pq,f, - P q , p , - l for the first eq - 1 levels of attribute q. The corresponding information matrix is still BhABL, given that all other interaction effects are assumed to be 0. We illustrate this in the next example. EXAMPLE 3.2.9.

Consider the situation of Example 3.2.5. Suppose that we assume that the two-factor interaction is 0. Then for the first representation

and hence

If we let

we can test hypotheses about the main effects using Bhy. For the second representation the parameters of interest are &,o - PLl, Pz,o - P 2 , 2 ,

P2,l - P2,2.


77

If I

1

1 -1 -1 -1 1 0 -11 0 1 - 1

0 -1 0 1 - 1

& = [ l

then

Supposethatweuse thepairs(00, l l ) , O l . 12),(02, lO),~lO,Ol).(ll,O2),and(12,OO). Then, under the null hypothesis, we have

A=-

1 4x6

2 0 0 0 - 1 - 1 0 2 0 - 1 0 - 1 0 0 2 - 1 - 1 0 0 - 1 - 1 2 0 0 -1 0 - 1 0 2 0 -1-1 0 0 0 2

and the information matrix for R ~ isYBc2ABL = Cn. The information matrix for BQY is often expressed in terms of the treatment combinations in each choice set (see Huber and Zwerina (1996), for example). We show the relationship between these two approaches here. For the second representation the x , are represented by triples; see Example 3.2.5. These triples are the columns of Bn and so we will write Bn = [ X I . x2, . . . , xfj

1.

Then

Huber and Zwerina (1996) write the information matrix for Bay,under the null hyQothesis, as

where there are ITL options in each of N choice sets and

In this case m

=

2. and hence we have

78


Since nZzJ= 1 when x, and x7are in the same choice set and is 0 otherwise we find

The other question is that of connectedness. Since we are no longer interested in the

t treatment combinations specifically but in the main effects (and perhaps the two-factor interactions), it is no longer necessary to use the definition of connectedness that we gave previously. Now it is only necessary for the effects that we want to estimate to be “connected”. This is harder to define combinatorially, but one practical way to see if a set of choice sets is connected for the effects of interest is simply to calculate C = Bh,AB&, and check that C is invertible. Certainly quite small designs can be useful, particularly if only the main effects are of interest, as the next example shows. EXAMPLE 3.2.10. If there are k = 3 binary attributes, then the pairs in Table 3.6 might be used as a choice experiment. (We say that an option has beenfoldedover if all the 0s i n an option have been replaced by 1 s and vice versa to get the other option in the choice set.) The corresponding A matrix, under the null hypothesis that all items are equally attractive, is given by

A=-

1

(1

1

+ 1)2 4

0

0 0 0

0 0

0

1 - 1 0 - 1 1 0 0 - 1 0 0 0 - 1 0 0 0 -1 0 0 0 0

0

0

0 0

0 1

0

0

0 1 0

0 0 1

1 0 0 0 0

0 1 0

0 0

0 1 0

0 - 1

0 - 1 0 - 1 0

0 0

0

Suppose that we are only interested in estimating main effects and that we can assume that the remaining contrasts of the yi are equal to 0. Then we have the three main effects contrasts in Bh and B, is empty. We find that C: = i I 3 , showing that we only need these four choice sets to estimate the main effects for three binary attributes if we can assume that the interaction effects arc 0. 0 In this section, we have parameterized the yi i n terms of the factorial effects, and have found the corresponding maximum likelihood estimates and variance-covariance matrices. We have seen that if Chr = 0, then BhT is independent of the form of B,. and that the definition of connectedness changes to reflect the effects that we want to estimate.

THE MNL MODEL FOR CHOICE SETS OF ANY SIZE

79

Table 3.6 The Fold-over Pairs with k = 3

Option B

Option A 000

111

110

00 1

101 01 I

010 100

In the next section, we derive the A matrix for largerchoice sets and briefly discuss using k attributes to describe the options in this situation. In Chapter 4 we discuss the structure of the optimal designs for binary attributes for the estimation of main effects, main effects plus two-factor interactions and main plus some two-factor interactions when all choice sets are of size 2. In Chapter 5 we allow the choice sets to have more than 2 options and in Chapter 6 we allow the attributes to have more than 2 levels each and the choice sets to be of any size. 3.2.6

Exercises

%‘jo) = tui/30 and perform 5 iterations. How do these estimates compare with the ones given in the text?

1 , In Example 3.2.2 use the initial estimates

2. Suppose that k = 4 and that the four attributes are binary. Suppose that the four pairs (0000, 1 1 1 I ) , (001 I , 1 loo), (0101, 1010), and (001 1, 1100) are to be used to estimate the main effects of the attributes. (a) Determine the matrix Bh for estimating main effects. (b) Determine the matrix A associated with these pairs. (c) Hence detemiine the information matrix C when B, contains all of the effects except the main effects and when B, is empty. Comment. (d) Find the !lo matrix for these choice sets.

3.3 THE MNL MODEL FOR CHOICE SETS OF ANY SIZE In this section we discuss the MNL model for choice sets of any size. Train (2003) gives an extensive discussion of other models and of associated estimation questions.

3.3.1 Choice Sets of Any Size We begin by deriving the information matrix for the 7r7 when the choice sets may have any number of options in them. Since all subjects see the same choice sets, we only need to consider results from one subject in the derivation; so the subscript cy has been suppressed in what follows. Consider an experiment i n which there are N choice sets of m options, of which nI,.i*.....?,,, compare the specific options T i l , 71, ~.. . ,Yi,,,, where 7’,,.?,.

,I,,,

=

{

1 0

if (Yil,Yi2 . . . ,Yi,,,) is a choice set. otherwise. ~

80


c

Then

N =

~,i1,z,,...,z"7

a1 < ? , 2 < , , . < i , , ,

We define parameters X = ( T ~ "2, , . . . , r t ) associated with t items Y;, Yi,.. . ,Ti.Given a choice set which contains m options Y;, , Y;,, . . . , Yin, the probability that option 71, is preferred to the other m - 1 options in the choice set is Ti1 P(Til > 21,:. . . , IZn, )= I

,

C,"=1 Ti7

for i j = 1 , 2 , . . . ,t (assuming that all options in each choice set are distinct). We also assume that choices made in one choice set do not affect choices made in any other choice set. If m = 2, this is just the model of the previous section. We use the method in Bradley (1955) and Pendergrass and Bradley (1960) to derive the form of the entries in A for any value of m. Let to,, ,iz ,...,i,, be an indicator variable where =

w i l , i z ,...,i,,,

{0

1 if '11, > Y;,, 71, . . . , 'li,,,, otherwise.

Then

We let wtl be the number of times option Ti,is preferred to the other options available in any choice set in which ?1, appears. Then Wil

=

C

t u i l r i 2 ,...,i n , ,

i l l:l=

n:,,;2 =

2 1

xuz;l=

xu,;2 =

1 1

x,,, :1 = 0 Xtl3:2 = 3

Table 5.4 All Possible Triples when k = 3 Sorted by Difference Vector vz = ( 1 , 2 , 3 ) (000,001.1 10) (000.001,111)

v3

(000,OlO.I01) (000,010,1 I I )

(000,101,110) (001,010,100)

(000.0l1,100) (000.0l1,1I1) (000.100.1 I I ) (000,101,11 I ) (000,110,111) (001,010.101) (001,010,110) (001,011,100) (001,011,110) (001,100,110) (001,101.1 10) (00 1, I 10,111) (010,011,100) (0l0,0I 1.101) (010,100.I01) (01 1,100.101) (010.101.110) (010,101.1 11) (011,100.110) (01 1.100,1l1)

(001,010.1 11)

= (2,2,2)

(000.0l1,101) (000.01 1,110)

(001.100,1 I I ) (010.100.111) (01 1.101,1l0)

5.1.2 The Derivation of the A Matrix In this section, we derive the A matrix when we assume that the class of competing designs consists of all designs in which all choice sets with a given difference vector are either all included in the choice experiment or none of them are. So, in Example 5.1 3,the competing designs are the choice sets from one or more of the columns of Table 5.4. Thus there are 7 competing designs to be considered in that case.

144

LARGER CHOICE SET SIZES FOR BINARY ATTRIBUTES

The matrix of contrasts is unaffected by the choice set size. However, we do need to derive the appropriate A matrix. From the results in Section 3.3.1, we know that, if we assume that = T 2 = .. . = R p = 1

(that is, all treatment combinations are equally attractive, the usual null hypothesis), then

(5.2)

and n,, r 2 2 , , indicates whether or not the choice experiment contains (Til, Ti,, . . . , I,,n ) as a choice set. Thus the diagonal entries of A are times the proportion of choice sets containing the treatment combination Ttl, and the off-diagonal entries are times the negative of the proportion of choice sets containing both T,, and 71,. In Section 4.1.1, the general form of the A matrix for m = 2 was given as a linear combination of the identity matrix of order 2k and the Dk,% matrices. We now derive a similar result for any value of m.

5

r ,

5

LEMMA 5.1.2. Under the usual null hypothesis,

where

z

= c c v , a v l=

-

3

and

The row and column sums of A are equal to 0.

Proof. We begin by counting the number of times each treatment combination appears in the design. The number of times that the treatment combination 0 0 . . . O appears in the choice cv, iv, , and this is the number of times that any treatment combination experiment is appears in the choice experiment because of the assumption that all choice sets with a given difference vector appear (or do not appear) in the choice experiment. (Throughout this section the summations over j are over all possible difference vectors v3 for the appropriate values of k and m.) So the proportion of choice sets in which any treatment

OPTIMAL DESIGNS FROM THE COMPLETE FACTORIAL

145

as required. Now we count the number ofpairs in the choice experiment that have i attributes different. If choice sets with difference vector v3 are in the design, then each treatment combination appears in c, choice sets with difference vector v3.and zv,;fof the differences are equal to i . Altogether, there are 'Lk treatment combinations and each treatment combination can appear in any of the m positions in the choice set. Thus the total number of pairs in the choice experiment with i attributes different is

2" ~ c v , i v , x v , TY1

2

(5.3)

J

Considering the 2k treatment combinations, the number of pairs with i attributes different is

Each of these pairs appears in the choice experiment the same number of times, say r,. Then the total number of pairs in the choice experiment with i attributes different can also be expressed as (5.4)

Equating (5.3) and (5.4), we get

So, if the proportion of choice sets in which each pair with 2 attributes different appears is y1 = r , / N ,then

Thus

3

from which we see that ?A =

2 (;) 1n

c,,

-

.I

Hence we have established that

a , ,2, ,: j .

146


Finally, we establish that the row and column sums of A are equal to 0. Summing over i , we get

0

as required.

WEXAMPLE 5.1.4. Let m = 3, k = 3, with v3,cV3 and xvJiias in Example 5.1.3. Assume that choice sets with difference vectors v2 and v3 only are in the choice experiment. Then i,, = 0 , i,, = i, = I , N = 24 8 = 32, a,, = 0, and avz = avs = 1/32. Thus we have

+

and

= -

=

2 -

1 1 x - (0 + 9 x - x 1 3 3 32 2 32 ’

2 3

1 3

- x -(0+9

1

1 x 0) +3 x 32

1 32

x - x 1 + 3 x - x 3)

32


147

2 1 1 - ( 0 + 9 x - x 1 + 3 x - x 0) 3 32 32 6 __ 32. Note that y? is the proportion of choice sets in which each pair with i attributes different appears in the choice experiment. For example, each pair with one attribute different appears in 2 of the 32 choice sets in the choice experiment: the pair 000 and 001 has one attribute different and appears in two choice sets, (000,001, I 10)and (OOO,OoI,l1 I ) , out of the 32 in the experiment Now that we have calculated the z and y, values, we can calculate the A matrix. For the A matrix we need the matrices. 0 3 , and ~ 0 3 . 2 are given in Example 4.1.2 and the 0 3 3 matrix consists of 1 s down the back diagonal and 0s in all the other positions. We get =

-

-

1

288

24 -2 -2 -4 -2 -4 -4 -6

-2 24 -4 -2 -4 -2 -6 -4

-2 -4 24 -2 -4 -6 -2 -4

-4 -2 -2 24 -6 -4 -4 -2

-2 -4 -4 -6 24 -2 -2 -4

-4 -2 -6 -4 -2 24 -4 -2

-4 -6 -2 -4 -2 -4 24 -2

-6 -4 -3;

-2 -4 -2 -2 24

In this section, we evaluate the information matrix for estimating main effects only and use this to determine the optimal choice experiment fork binary attributes using choice sets of size ni. As in Section 4.1.3 we let B2r be the contrast matrix for main effects and we let CAf be the k x k principal minor of C = B2k AB;, associated with the main effects. Thus we are evaluating the information matrix when Bh = Bzk , A r . In Lemma 4. I .3 we have shown that k-1 k-1 R2k.AlDk.i = i-1 Bzb,Af

[( ) ( )]

for all allowable i. We now show that at this stage we do not need to make any assumptions about the contrasts in B,. since the class of competing designs that we have chosen ensures that Ch,. = 0 for any choice of R,. By definition,

148


Since Bh = B 2 k , A - I we can see that

- 0k,2F-k-a.

This is true for any choice of B, as long as the set of competing designs remains the same so that the form of A stays as a linear combination of the D k , z . Using the result in Lemma 4.1.3,we now can get an explicit expression for the information matrix, Chi, for the k main effects.

H LEMMA 5.1.3. The infomation matrix for main effects under the null hypothesis is given by

Proof The information matrix for main effects under the null hypothesis is

CM

=

B~~,A.IAB~~,A~

Since

as required.

Thus the determinant of C ~ . i s


149

w EXAMPLE 5.1.5. Let m = 3 and k

3. There are two ways that we can calculate CAI: by evaluating given i n Lemma 5.1.3. In this example we will do both. The matrix B2i nr is rhe contrast matrix for main effects and is given by =

CAI = B ~nrABi,,,, J or by using the expression for

-1

-1

-1 -1

-1 -1 1 1 1 1 -1 -1 1 -1 1 -1 1 -1

-1

1

In this example, we assume that all choice sets with difference vectors v2 or v3 are in the choice experiment. Then we can use the A matrix from Example 5.1.4 to calculate 1

= 913:

CAI= B 2 ’n1ABaJ

thus det(C41) = (;)3

Alternatively, we can use Lemma 5.1 3 to calculate C,ql Then, using the y1 values given in Example 5.1.4,

8 1 0 9 To find the D-optimal design, we must maximize det(CA1) subject to the constraint 2”/m = 1 (since the total number of choice sets must equal AV). 1 and using the fact that (:I;)/(:) = z / k Substituting gJz = gives =

-13.

?(:)- 1,

df?t(CAI

The following theorem establishes that the D-optimal design, for estimating main effects only, is one which consists of choice sets i n which the sum of the differences attains the maximum value given in Lemma 5.1. I.

w THEOREM 5.1.1. The D-optimal designfor testing main effects only, when all other effects are a.ssitmed to be zero. is given by choice sets in which, ,for each v,j present, 111

c

( m - 1) / 2 ,=1

”I,

(m2- l ) k / 4 .

111

m2k/4,

in cveii,

odd.

=

150

LARGER CHOICE SET SIZES FOR BINARYATTRIBUTES

and there is at least one vI with a non-zero av3;that is, the choice experiment is noiz-empty. For the optimal designs

Proof Recall that av7 =

-zv,_ -

iv, 2k c h cVir iV,

and that i, is an indicator variable indicating the difference vectors corresponding to the choice sets that are included in the choice experiment. Thus at least one of these i V 7 and , therefore av,, values must be non-zero. Substituting for uv, in det(CA[), we have

Since xv,; i denotes the number of times the difference i appears in the difference vector vg,multiplying by i and summing these is equivalent to summing the entries in v3; thus,

(y)

k i=l

1=1

Therefore

m(m-l)/2

d,, can When rn is even, Lemma 5.1.1 states that the maximum value that Cz=l attain is m 2 k / 4 . Then it follows that CI"_'y-1'/2 d,, = rn2k/4 - p, for some p , 2 0. Thus

=

[-1

2k

c,

PJCV, iv,

r n 2 k 2 k - 2 xh cvl,iv,

The cv, values are all positive and the i,, are either 0 or 1 ; so, for rn even, det(C~4)has a maximum of ( 1 / 2 k ) k when p , = 0 for all 2. Thus we obtain the maximum det(CAr) when

C i=l

d,, = r n 2 k / 4


151

Similarly when m is odd, Lemma 5.1 .I states that the maximum value of ~ ~ ' ~ - ' )d2' 7 ' is ( 7 1 7 ~ 1)k/4. Then d,, = (in' - l ) k / 4 - p , forsomep, 2 0. Substituting, we have

cE(r-1)/2

Since the cv, values are all positive and the i,, are either 0 or 1 , det(CA1) has a maximum of ((771' - 1)/(m22k))kwhen p , = 0 for all 3. For m odd, we obtain the maximum det(Cn1) when

c

m(m-l)/Z

d,, = (7n'

-

1)k/4.

1=1

Therefore, the maximum value of det(C,ff)is

This occurs when

c

( m 2- l)k/4, m odd,

m(m-1)/2

4, =

1=1 m2k/4, m even. The D-efficiency relative to the optimal design is calculated using the expression

where p = X:, the number of main effects we estimate. We now look at all possible designs for a small example. EXAMPLE 5.1.6. Recall that, for V L = 3 and k = 3 , there are three difference vectors: v1 = (1.1.2), v2 = ( 1 , 2 . 3 ) and v g := ( 2 , 2 . 2 ) . All the possible designs for choice experiments in this situation are given in Table 5.5. Since m, is odd, Theorem 5.1.1 states that the Doptimal designs have choice sets in which the entries in the difference vectors sum to ( m 2 - l)X:/4 = 'Lk = 6 and a maximum det(Copt.nl)= (l/9)3. The difference vectors VP and v3 have entries which sum to 6; so there are three D-optimal designs:

1 . All 24 triples with difference vector v'. So a,? = 1/24 and a,, = n,, = 0. 2. All 8 triples with difference vector v3. So a,,,= 1/8 and a,, = a,,

= 0.

3. All 32 triples with difference vectors v2 and v3. So (rvz = a,,,3 = 1/32 and n,, = 0. The smallest of these D-optimal designs is the second one, consisting of the following eight triples, each with difference vector v3: (000,011; 101). (000,011,110), (000,101,110), (001,010; 100). (001.010,111). (00l,100,111)~(010,100.111). (011.101.110). 0

152


All Possible Choice Experiment Designs for Binary Attributes when k = 3 and

Table 5.5 m=3 Vi

Ei&j

in Design

4 6 6 4,6 4,6 6 4, 6

v1

v2

v3

Vl

&vz

v1 & v 3

vz

&v3

v1, v2

& v3

i,

i,

i,

1 0

0 1

0 0

0 1 1 0 1

0 1 0 1 1

1 0 1 1 1

N

24 24 8 48 32 32 56

a,, a,,

av3

Efficiency (%)

0

66.67

O h 0

100

L 4p L 48 :

100 83.33

z O ! - + -

15 100

k

0

o o ? 1 56

3: 56

56

85.71

5.1.4 The Model for Main Effects and Two-Factor Interactions To find the D-optimal designs for estimating main effects and two-factor interactions for any choice set size m, we generalize the results of Section 4.1.4. As in Section 4.1.4, we let B2k,,Mbe the rows of B2k that correspond to main effects and we let B2k,T be the rows of B2k that correspond to the two-factor interactions. The matrix associated with main effects and two-factor interactions is denoted by and is the concatenation of B2k and B2k,T.For the D-optimal design we evaluate the

[ + (k)] [ + (i)]

k x k principal minor of C = B2k As we are working with the complete factorial, it is not important how the contrasts for interactions of more than two factors are divided between B, and Bh. In Lemma 4.1.5. we established that B2k,TDkt=

[(k - 2

k-2

k-2

)-2(i-I)+(i-2)]B2k.T

for all allowable i. LEMMA 5.1.4.

Under the null hypothesis, the information matrixfor main effects plus two-factor interactions is given by

Proof. We let Bh = B 2 k , A f T and then


153

Now

As we noted previously,

so

(r Ii)

and hence

(f)

= ("1)

=

(;)

(! 1;) + (:I :)

-

("2)

+(;I;): and ( k

i

I)= ( k

i' )+ (k - 2l ) .

(;I;) =(;I;).

+2(fI?) -

and we have

0

as required.

Hence the determinant of

(7.41~

is

EXAMPLE 5.1.7. = 3 and k = 3. We can calculate CAITeither by evaluating

Let in

CI?IT =

BZJ.AlTAB43 A l T '

154


or by using Lemma 5.1.4. In this example, we will do both. The matrix B 2 3 , M T is the contrast matrix for main effects plus all two-factor interactions and is given by

1 B Z ~ , A=I T2 8

-1 -1 -1 1 1 1

-1 -1 1 1 -1 -1

-1 -1 1 1 1 -1 -1 -1 1 -1 1 -1

1 1 1 -1 -1 1 1 -1 -1 1 -1 -1 1 -1 -1 1 -1 -1

1 1 1 1 1 1

In this example, we assume that all choice sets with difference vectors vz or v3 are in the choice experiment. Then we can use the A matrix from Example 5.1.4 to calculate

We can also use Lemma 5.1.4 to calculate CbfT. From Example 5.1 .S,

Then, using the yi values given in Example 5.1.4,

=

1 12

-13.

Hence

To find the D-optimal design, we need to maximize d e t ( c , t f T ) ,subject to the constraint that 2%/m = 1. THEOREM 5.1.2.

The D-optimal design for testing main effects and two-factor interactions, when all other efferts are assumed to be zero, is given by designs where

T m(m-1) ( ~ ,k+l~ ) - ' ,

k e v e n a n d i = k/2, k / 2 + 1 ,

w((k$),z)-l, + koddandi = ( k

0,

otherwise,

1)/2,


155

when thiy results in non-zetv y7 valuev that correspond to difference vector7 that artually eiirt. The niaximum \ulue of the determinant is then

PmoJ In Theorem 4.1.3, we proved that f o r m = 2 the D-optimal design for testing main effects and two-factor interactions is given by k even and?, = k / 2 . k / 2

{2k-'($i)}-1

k odd and {2k-'((A.4),2)}

L'

= (k

+ 1.

+ 1)/2.

otherwise.

0

In the proof of this theorem, the function f = W 2 ( k - 1 )and / 2 ,therefore det ( ( ' I Y I T ) , 15 maximbed subject to the constraint C,"=, (!)xu = 1. where

and k

k-2

v=l

For choice sets of size vi the constraint is

In order to use the results in the proof of Theorem 4.1.3, we let 2, =

(suggested by Moore (2001) when a$ we had for 7n = 2. Now we let

and

7n

2'gyz/(m(?n- I ) )

= 3). Then we have the same form of the constraint

156


zf=,

is maximized subject to the constraint (t)z, = 1 for the same z, values given above for the m = 2 case. Using z, = 2 k y t / ( m ( m - l)),we obtain the optimal designs in terms of the yz values, as required. We now determine the maximum value of the determinant at these y, values. For k even, only two values of i will give the maximum determinant. These are i = k / 2 and i = k/2 1, where

+

yk/2

= Yk/2+1 =

2k

and all other yz = 0. Then r

i k

r

Therefore, for k even, det(CAfT) =

(

( m - 'Itk i'I) m(k 1)2k

k+k(k-l)/Z

+

For Ic odd, the only value of i that will give the maximum determinant is i = ( k where k Y(k+1)/2 = ') ((k + 1112)

;:"'

and all other y, = 0. Then

+ 1)/2,


157

Therefore, for k odd,

Hence the maximum value of the determinant is

Note that for any design we can calculate the D-efficiency relative to the optimal design using the expression

+

where p = k k ( k - 1)/2, the total number of main effects and two-factor interactions that we are estimating. The following two examples illustrate the use of Theorem 5.1.2 to obtain choice sets with the maximum value of the determinant of C;tlr. EXAMPLE 5.1.8. For m = 3 and k = 3, Theorem 5.1.2 states that the D-optimal design is given by

I O.

otherwise.

The maximum value of [let ( C J ~ Tis)

Since 4av, + 2 n v , = 0,

y1

=

y2

=

2av,

+ 2nv, + 2av3 = +.

y3

=

6lv2

= 0,

we see that av, = a,vn := 0 and nvs = $. Thus the D-optimal design consists of the 8 triples with difference vector v3 = ( 2 , 2 , 2 ) . Information about the 7 competing designs is shown in Table 5.6. 0 EXAMPLE 5.1.9. 7~ = 3 and k = 4 the possible difference vectors are v1 = (1,1,a), v 2 = (1,2,:3), v g = ( 1 , 3 . 4 ) , v4 = (2.2.2), v 5 = (2.2,4), and v g = ( 2 . 3 . 3 ) . Using Theorem 5.1.2,

For

the D-optimal design is given by

LARGER CHOICE SET SIZESFOR BINARY ATTRIBUTES

158

Table 5.6 The Efficiency, for the Estimation of Main Effects plus Two-Factor Interactions, of the 7 Competing Designs when k = 3 and m = 3

The maximum value of det(CA,rT)iS

Since yz and y3 are the only yEvalues that are non-zero and v4 and VG are the only difference vectors containing all 2s, all 3s,or a combination of 2s and 3s, we see that a,? = 0 unless j = 4 , 6 . Thus we have

YZ

= =

y3

=

+ 6av, + 2av, = 0, + 4av2 + 4av4 + 2av5 + 2av6 = 6/160, 6av2 + 2a,, + 6av, = 6/160,

y4

=

8ava +6a,,

y1

6avl 2avl

= 0.

The solution is a,, = avz = av3= ava = 0 and avr = a,, = 1/160. Thus the D-optimal design consists of the 64 triples with difference vector v 4 = (2,2,2) and the 96 triples with difference vector v6 = ( 2 , 3 , 3 ) . If we use just the 64 triples with difference vector v4, then this design is 99.03% efficient, and for the 96 triples with difference vector v g only, the design is 99.60% efficient. 0 However, for some values of m and k , solutions to the yz equations do not exist. For example, when m = 3 and k =. 1 (mod 4), no solution exists; the following example illustrates this case. EXAMPLE 5.1.10. If we let m = 3 and k = 5 , then the D-optimal design given by Theorem 51.2 is

y (0 -1

y3 =

>

y1 = yz = y4 = y5 = 0,

This means that triples with difference vector (3,3,3) are required since any other triple would result in one of the other gi values being non-zero. However, the only possible difference vectors are ~1

= (1.1,2), vz = ( 1 , 2 , 3 ) , ~3 = (1,3,4), ~4 = (1,4,5), ~5 = (2,2,2),

SMALL OPTIMAL AND NEAR-OPTIMAL DESIGNS FOR LARGER CHOICE SET SIZES

vf,

(2, 2.4). v7 = ( 2 . 3 .:3).

Vg

= (2,3.5), V9 =

(2,4,4), V10

159

= (3.3.4):

so no triple with the difference vector ( 3 , 3 , 3 )exists. By checking all possible designs, it can easily be shown that the optimal design consists of the 960 triples with difference vector vi = ( 2 . 3 , 3 ) and the 480 triples with difference vector vl0 = ( 3 . 3 . 4 ) , where y? = 1/1440. = 9/1440, y4 = 3/1440 and y1 = y,5 = 0. Thus avr = a,,,, = ] / I 4 4 0 and nvi = 0 for .j = 1.2.3.4.5,6,8.9.The maximum obtainable determinant is

As Example 5. I . 10 shows, the optimal designs derived in this and the preceding sections can become very large as the number of attributes increases. The question of how many choice sets can be included in a stated choice experiment has been considered by various authors. Braze11 and Louviere (1 995) show that choice experiments with up to 128 choice sets can be effective in parameter estimation. In the next section, we investigate the D-efficiency of small designs obtained from a generalization of the constant difference construction in Section 4.2.3. 5.1.5

Exercises

1. Let k = 3 and m = 4. Give the 35 sets of 4 treatment combinations with 000. Calculate the difference vector for each. Hence calculate the cy, and, for each vJ, calculate the z v , : i Verify . that the bound of Lemma 5.1 . I is correct.

2. Use Lemma 5. I . 1 when k = 3 and m in this case?

=

3. Let k = 3 and m == 4. Evaluate

yz and y3. Hence give an expression for A.

2 , y1,

4. Can you get more than one set of rows

4. I f i n = 2 show that yz =

5. Let k = 3 and rn == 4. Using Theorem 5.1 . I give all the possible D-optimal designs for estimating main effects. Can you use a smaller subset of any of these to get a smaller design which is still D-optimal‘? 6. Let k = 3 and m = 4. Use Theorem 5.1.2 to give the v, that would correspond to the D-optimal designs for estimating main effects plus two-factor interactions. Do the y, correspond lo an actual design? If so, give it. If not. try 2 or 3 designs “close to” the optimal p 1 and compare them. Which would you recommend?

5.2


In this section, we give constructions for small optimal and near-optimal designs for choice sets of size m. The results are an extension of results in Section 4.2, where we gave constructions for optimal and near-optimal designs for estimating main effects only, and for estimating main effects plus two-factor interactions. Recall that the constructions there started with a fraction of resolution 3 (for estimating main effects only) or resolution 5 (for estimating main effects plus two-factor interactions). Pairs were formed by adding

160


one or more generators to the treatment combinations in the fraction, where the addition is performed component-wise modulo 2. Each generator gave rise to a set of pairs. We had to assume that all contrasts other than the ones we were estimating were 0 however. So B, contains all the contrasts that not in Bh and B,. is empty. 5.2.1

The Model for Main Effects Only

When estimating main effects only, for each attribute there must be at least one generator with a 1 in the corresponding position. Using one generator consisting of all 1 s, which is equivalent to using the resolution 3 fraction and its foldover, results in a D-optimal design in a minimum number of pairs, in which all main effects can be estimated. The next result gives generators for small optimal designs for estimating main effects when m 2 2.

THEOREM 5.2.1. Let F be a fractional factorial design of resolution at least 3 and with k factors. Let G be a .set of generators G = (gl,g2, . . . ,gm),where each g , is a binary k-tuple, and gl = 0. kt v = ( d l , d2,. . . , d m ( m - l ) / 2 )be the differenre vector consisting of all the pairwise differences between the generators in G, where m(m-1)/2

(m2 l ) k / 4 , { m2k/4, -

i=l

di =

+

m odd, m even.

+

Then the choice sets given by ( F ,F gp, . . . , F g m ) , where the addition is done component-wise modulo 2, ore optimal for estimating main effects only. Pro06 Let IF1 = 2 k - p and let BF be the columns of Bhf corresponding to the treatment combinations in F . Thus BpB; = T I k . Let D,, be the diagonal matrix with an entry of 1 in position j if the j t h entry of g , is 0 and an entry of -1 in position j if the jth entry of g , is 1 . Then D,, RF has the same columns as the columns of Rnf for the treatment combinations in F g , . We construct the choice sets from F by adding the generators to F , using modulo 2 arithmetic. Then the choice sets are ( F ,F g z , . . . F g m ) , where each row represents a choice set. Let N be the number of choice sets of size m (hence N = 2 k - p if there are no repeated choice sets) and let n,, be the number of choice sets that contain the pair of treatment combinations u,w. This number may be 0 if at least one of PL and w does not appear in thechoice sets. Let n,,,,d be the number of choice sets which contain u and PI in columns c and d (unordered). Then nu, = n,,,,d, and we define nu, = nuv.Then the values nu, and n,,, are the entries of m2N A . Consider two columns c and d . Then we know that column c contains the treatment combinations in F + g, (in some order) and column d contains the treatment combinations in F + gd in some order. I f F + g, = F g d , then g , gd E F and we can write F as F1 U (F1 g , g d ) . Then, using this order for the elements of F , we have that the submatrix of A corresponding to the elements in F can be written as (reordering rowq and columns if necessary)

+

+

+

xu+,

xc,d

+ +

+

+


161

since we know that each treatment combination in F gives rise to a distinct choice set. We can write R,jr as Bnr = [ R F , D g , + g , , H ~ B I F ] .Now

so

If g ,

+ g d $ F , then the same argument establishes that

Now we know that m w-1)/2

7=1

d’ =

{

( m 2- l ) k / 4 . m2k/4.

m odd. in even.

Consider the contribution to the C ,d, in a choice set from just one attribute. The attribute will have T 0s and ( r n - z) 1 s to give a total of z ( m - T ) . This contribution is maximized by having m / Z Is and ? n / 2 0s for 172 even, and ( m- 1)/2 0s and ( m 1)/2 1 s (or vice versa) for I J J odd. Thus the maximum contribution to C ,d, from any one attribute is m2/4 for 772 even and ( m 2- 1 ) / 4 for m odd. Thus we see that each attribute must contribute exactly this amount to C7d,for the optimal designs. suppose 7n = 21711 1. NOW D ~+g,, , is summing (””+’) entries of 1 in each diagonal position and m l ( 7 n l + 1) entries of -1 in each diagonal poGtion. Thus, if rn is odd.

+

+

c ~ , ~

(71)+

Similarly, if ?nis even,

+

Suppose that none of the g, is in F and that g, g d $ F for any pair c and rl. Let A be the set of treatment combinations that appear in the choice experiment. Then

162


Thus we have

as required. Now suppose that, for at least one pair of columns, the number of pairs is 2 k - p - 1 (that is, F g , = F gd for some c, d). Then we see that

+

+

2k-P

NBhZhcdB61 =

T ( I k - Dg,+g,,).

Although this works out for the pairs, once the pairs are considered as part of the larger choice sets, then the pairs will i n fact appear twice and so we need to use 2k-p+l

2NBnlA,dBb

=

(Ik 2k

- Dg,A

~~~~~~~~~

0

~

~

0 0

0

0

~

0

s

0

o

0

0

& o o o o . &

o

o 0 0

o 0 0

o 0 0

o 0 0

0 0

o 0 0

0

0

0 0 0 0

0 0 0

0

0 0 0 0 0

0

0

0

0

0

0

0

0

.& o ~0

o 0

1 2 3

0 0 0 0 0

0

o

32

& o . & 0 0

0

ft

G o

0

We see that det(chfT) = 0 and so no effects can be estimated. If we now assume that each choice set has a none option adjoined, then a row and column is ad,joined to C A ~ for T the none option and we have

as we would expect from Theorem 7.1.1. Thus we see that the matrix B j D B ; is not a diagonal matrix and in this case adjoining the none option has not improved the properties of the design since we have det(Ch{T.,) = 0. 0 The result in Theorem 7.1.1 holds for any stated preference choice experiment in which a none option has been adjoined to each choice set. Obviously the expression for C, is easier to work with if B j r = 0,. This is true if the treatment combinations that appear

OPTIMAL STATED CHOICE EXPERIMENTS WHEN ALL CHOICE SETS CONTAIN A SPECIFIC OPTION

Table 7.2 Choice Sets with k = 3 Attributes

Option 1 000 I10 I l l

00 I 012 102 103 013

(1

233

= Pz = 2 and 83 = 4

Option 2 111

00 1 002 I12 103 013 010 100

in the choice experiment form a fractional factorial design in which each level of each attribute is equally represented. 1 3 ~ I l R ;will then be a multiple of the identity matrix if the treatment combinations form a regular symmetric fractional factorial design. In this case effects which are not independently estimated in the original forced choice experiment will not be independently estimated when the none option is djoined but sometimes effects which can not be estimated in the original forced choice can be estimated in the extended design as we have seen for the complete factorial i n Example 7. I . I .

7.1.2 Optimal Binary Response Experiments In a binary response design respondents are shown a set o f options, in turn, and asked for each option whether or not they would choose it. Thus this is really the simplest example of having a none option i n each choice set. The options themselves might be actual products for sale or they might he treatments i n a medical setting, for instance. Hall et at. (2002) used a binary response experiments to investigate chickenpox vaccination. In the next result we prove that showing the products determined by a resolution 3 fraction results in a diagonal C matrix with the largest possible determinant amongst designs with a diagonal (’ matrix. Thus each choice set is of size 2: the product under investigation and the “no” option. THEOREM 7.1.2. Sirppose thot the mnin effect,$of the nttrihiites for options descsrihed by A: nttribirtcs, where the qth attrihiire /ins f , levels3 are to be estimated m i n g n 17inar.y re.yporzse design. 7’hen n re,~olution3, eqiii-replicnte,fractiongives rise to n diagonal infonnntion mntrix~forthe rstiriintion of mnin elfects ond it 110.~the largest determinnnt nniong.ct de,~igri,t with dingono1 infoomintion mntrice.s. ~ ‘ / ? L I . F

Proof Let us assume that A’ of the L products are to be presented to respondents. Then we can order the products so that these N products are the first N products in some fixed, hut arbitrary, ordering of the L products. We use this ordering to label the first 1, rows and columns of A and the first L columns of B. The final row and column of A and the final column of B correspond to the “no” option. Thus we obtain the A,, and B7,matrices as we did in the previous section. Indeed R,, is exactly the same as i t was in Theorem 7.1 . I .

234

VAROUS TOPICS

Since each product is shown individually to each respondent the number of choice sets is N and we can write

We then get that

which is a special case of the result in Theorem 7.1 . I .

0

7.1.3 Common Base Option If a particular combination of attribute levels appears in all choice sets then this combination is called the common base option. It may represent the current situation or the current treatment for a particular health condition (examples in Ryan and Hughes (1997), Ryan and Farrar (2000),and Longworth et al. (2001)) or the common base may be randomly chosen from the main effects plan and have all the other scenarios from the plan compared to it pairwise (examples in Ryan (1999), Ryan et al. (2000),and Scott (2002)). Our first optimality result gives the determinant of the information matrix for any resolution 3 fractional factorial design in which any one of the treatment combinations may be used as the common base and only one other option appears in each choice set. To ensure that the matrix of contrasts for main effects for the treatment combinations that appear in the choice experiment is unambiguously defined, we will insist that the treatment combinations that appear in the choice experiment form a fractional factorial design of resolution 3. From Theorem 6.3.1, we know that the maximum possible determinant for the information matrix for the estimation of main effects only, when choice sets are of size 2, is given by

We will use Cc for the information matrix when there is a common base. THEOREM 7.1.3. Let F be a resolution 3 fractional factorial design with equal replication of levels within each attribute and with N treatment combinations in total. Choose any treatment combination in F to be the common base. Then

The optimal design arises from the smallest resolution 3 design and so

Pro05 Without loss of generality, order the treatment combinations so that the first treatment combination is the common base, the next N - 1 treatment combinations are the other

OPTIMAL STATED CHOICE EXPERIMENTS WHEN ALL CHOICE SETS CONTAIN A SPECIFIC OPTION

-

4 ( N - l)A,

- (N-1) -1 -1

=

-1 -1 1 o

0

1

.

” ‘

-1

-

0 0 1

OL-N,N

- - (N-1) -1 -

1 [Bl B21 4(A: - 1)

-

1 BI 4 ( N - 1)

0 o... 0L - N.N -1 . . . -1 -

o...

0

-1

0

l . . .

O

0

. ... 0 ...

1

-1 -1

4(1V - 1 ) B1

-1

0 0

-1

-1

0

(Recall that det(ITl+ J p ) = ( p + I).)

1

OL-N.L-K

-

R;

. . . -1 -

o...

0 ... . ...

-

...

1

- (AT-2)

1

-1

[

O,\F,L-K

’..

-1

-1

-

-1 -1 . . . -1 1 0 ... 0 ,

- (N-1)

-

OL-N.L-.V

-1 -1

0,V.L-n.

...

o...

0

-1

...

...

235

o...

0 0 .

1

B:+*(N-I)

B*q

0-

0

The efficiency of the design, relative to the optimal forced choice stated preference experiment with choice sets of size 2, is given by

Using Theorem 7.1.3, we can see that, from the point of view of statistical efficiency, all resolution 3 designs with the same number of treatment combinations are equally good for

236

VARIOUS TOPICS

a particular number of attributes with given number of levels. It is also immaterial which of the treatment combinations is used as the common base. EXAMPLE 7.1.3.

Consider an experiment where there are four attributes describing each option. Three of the attributes have 2 levels and one has 4 levels. The smallest fractional factorial design of resolution 3 has 8 treatment combinations. Then using any 2 x 2 x 2 x 4//16 and using any treatment combination as a common base gives a design which is 93.3% efficient relative to a design with 8 treatment combinations and 45.17%efficient relative to a forced choice stated preference design. Using the complete factorial and using any treatment combination as a common base gives a design which is 90.3% efficient relative to a design with 8 treatment combinations and 43.71% efficient relative to a forced choice stated preference design. Using a 2 x 2 x 2 x 4//8 gives an efficiency of 48.40% efficient relative to a forced choice stated preference design. 0 In Scott (2002) a similar experiment was carried out with four attributes (two with 2 levels, one with 3 levels, and one with 4 levels) and using 16 choice sets. In this case the efficiency depends on the treatment combination chosen to be the common base. If the common base has the level of the 3-level attribute which appears 8 times, then the efficiency is 44.49%; and if the common base has either of the other levels of the 3-level attribute, then the efficiency is 46.12%. If we want to use this approach to estimate main effects plus two-factor interactions, then we need to start with a resolution 5 fractional factorial design. Unfortunately, in this setting we can only know that all effects are estimable and compare particular designs. We cannot calculate the efficiency relative to the optimal design since only a general expression for det(C) is available; see Chapter 6.

7.1.4

Common Base and None Option

The earlier results can be combined to give an expression for the information matrix for choice sets when there is both a common base and a none option in each choice set. Although in theory the choice sets could be of any size, we will only consider the situation when the choice sets are of size m = 3 and contain the common base, the neither of these option and one other option which will be different for each choice set. In this case the correct contrast matrix is B,, defined in Theorem 7.1 .I. By re-ordering the treatment combinations if necessary, and assuming there are N treatment combinations

Acn

=

1 9(N - 1)

-2(N-1) -1 -1

-1 2 0

-1 0

0 0

0 0 - ( N - 1) -1

-1 0 2

...

0

...

...

... .. . . . .

o...

.. . . . .

o... -1

-1 0 0 2 0

0 . . . -1

0 0 0

. . . 0 -(N-1)

... 0

... 0 .. . . .. . . . 0 ... 0

o . . . o

... . .. ...

0 . . . 0 0 ... 0

-1 -1

-1 0

0

2 ( N - 1)

If we assume that the treatment combinations involved in the experiment form a fractional factorial design, it is straightforward to show that the information matrix when there is a

OPTIMAL CHOICE SET SIZE

237

common base and a none option is Ccn,given by

where b, is the first column of Bnr.

7.2 OPTIMAL CHOICE SET SIZE In the previous chapters in the book we have assumed that we want optimal designs for a particular choice set size m. In this section we show that the statistically optimal choice set size is a function of the number of levels of each of the attributes. We find that although pairs are the most common choice set size used they are rarely the most efficient size to use. Whether the choice sets in these designs may prove to be too large for respondents to manage comfortably is something that potential users need to consider. The results that we give can be used to determine the relative efficiency of practical values of 71).

7.2.1

Main Effects Only for Asymmetric Attributes

In this section we consider choice experiments in which each of the attributes can have any number of levels and in which main effects only are to be estimated. From the results in Chapter 6 we know the maximum value of the determinant of CAIfor a particular value of r n . This allows the efficiency of any design to be calculated for particular values of ?n,k andt,,q= l , . . . > k . The following theorem establishes an upper bound for the maximum value of det(Cn1) when m is not fixed and gives the values of m for which this bound is attained. THEOREM 7.2.1.

Forfixed i~alrresof I; and F,, q = 1 , . . . . k , when estimating main efects o n l ~the . upper hound,for the maximum valire qf!fdet(CAr) over all valires of m is

k where p = C,=l (P4 - 1). The irpper boirnd i s attained when ni is a miiltiple of the lemt common mitltiple of I;, , P2. . . . , P,.

ProoJ: Let det,(COPt,,jr),be the maximum possible contribution of attribute q towards det (Copt, 4 I .any 1. CI early k

~ e t ( ~ o , t , , ~ ~7 n, )a= n,

(let((JO,t.nr),.

,=l

so we can examine det(CO,,t.nr),for each attribute separately. For each attribute q there are four possible cases to consider. 1 . Suppose m = cP,, where c is a positive integer and &, 5 m. If kq = 2 , then m is even and S, = nb2/4.Then for attribute q

238

VARIOUS TOPICS

If e,

+ y, where x = c and y = 0. Thus S, = (m2- ( ~+ 2xy ~ +2y ) ) / 2 = c2t,(t,

# 2 , then m = t,x

- I)/Z

and for attribute q

2. Suppose m # c&,, where c is a positive integer and 2 < t, < m. Then we have S, = (m2- (&,z2 + 2xy + y))/2 for 0 < y < C, and the determinant for that attribute is given by

We need to show that this determinant is less than the upper bound. That is, we need to show that

or

+ + y)] - m2(e,- 1) < 0. This is true because tq(C,x2 + 2zy + y) > m2 for all possible values of C,, [e,m2- t,(t,z2 2x21

z, and

y. Thus the determinant for attribute q in this case will be less than the upper bound.

3. Suppose that 2 5 m < C,. In this case S, = m(m- 1 ) / 2 and therefore

since ( m- l ) / m< (C,

-

l)/Cq.

4. Suppose t , = 2 and m is odd. Now S, = (m2- 1)/4 and for attribute q

This determinant is less than the upper bound since (m2- l ) / m 2 < 1. In cases 2,3, and 4 the maximum value of det(Chf)is always less than the upper bound, which is only ever attained for case 1 attributes. Thus the upper bound can only be realized 0 if all attributes belong to case 1. In the following examples we determine det(Copt,A.i,m)for values of m from 2 to 16. The D-efficiency for each of these values of m is calculated relative to the maximum determinant over all values of m, det(Copt,M);that is,


239

EXAMPLE 7.2.1.

Suppose there are five attributes ( k = S), each with four levels (P4 = 4 for q = 1 , . . . 5) and that the choice sets each have four options (m = 4). From Theorem 7.2.1 det(C'oL,t..~~.any r n ) is (1/4'))'; which is attained when 7n is a multiple o f 4 . In Table 7.3 the determinants and efficiencies are given for different values of In. Note that the choice set size of I n = 4 is optimal, as is the choice set size of ITL = 8 and other multiples of 4, 0 and that pairs ( 7 n = 2) are not nearly as efficient as other values of m. Table 7.3 Efficiency of Different Values of m for Five Attributes, Each with Four Levels for Main Effects Only.

m

det ( Copt.A T , )

Efficiency (%)

2 3 4 5 6 7

1.60004 x 1.19733 x lopd6 7.00649 x lop4' 3.79812 x 3.97781 x 5.142.54 x 7.00639 x 5.81534 x 5.72872 x 6.18641 x 7.00649 x 6.40972 x 6.32461 x 6.55365 7.00649 x

66.67 88.89 100 96.00 96.30 97.96

8 9 I0 11 12 13 14 15

16

100

98.77 98.67 99.17 100

99.4 I 99.32 99% 100

The following example illustrates using the theorem to choose values of m . EXAMPLE 7.2.2.

Consider four attributes where PI = 2, P2 = 3, and P3 = I s = 4. The least common multiple lcm(2.3,4) is 12 and this is the smallest value of ni for which det(C,'ol,+.,~~.anv ,,,) is attained. When I T ) = 12.

I

det(Copt.Af.any m ) =

[i]

z;=*(?#/-l) 9

[I]

=

= 1.44397 x lo-''

since all four attributes belong to case 1. Now choice sets of size 12 are too large for respondents in practice but knowing the efficiency of other values of 177 relative to this upper bound is useful when deciding what size the choice sets should be. Suppose ~n = 3 . For 41 = 2 and ~ nodd . (case 4),

For 8 2

=3

(case I ) ,

240

VARIOUS TOPICS

and for 3! = & = 4 (case 2),

Hence

det>(CAf)=

(A) ( (

(A)

3

&)3

&)2

= 6.33128 x lo-’’

The efficiency compared with using m = 12 is (6.33128 x lo-’’) 1.44397 x 10-18

I”

x 100 = 91.25%.

Now consider m = 4. Attributes 1, 3, and 4 all belong to case 1 and will have determinants equal to the upper bound, while attribute 2 belongs to case 3 since 2 < e2 < m. Now m, = e2x y so x = y = 1 and for attribute 2

+

Thus

(A) (A)(&) (&)

3

d e t ( C o p t , ~ ~= .4)

= 1.26912 x

with an efficiency of 98.58%. Table 7.4 gives the maximum determinant and the efficiency for values of m from 2 to 16. Observe that as long as ni > 2 the efficiencies are greater than 90%. 0 7.2.2

Main Effects and Two-Factor Interactions for Binary Attributes

In Chapter 5 the upper bound for the determinant of the information matrix, C A ~isTestablished, when estimating main effects and all two-factor interactions, when all k attributes are binary. The maximum possible value of det(Ch1T) for fixed values of m and k is given bv

Clearly, as m --t m, ( m - l ) / m of det(C:nfT),which is

--t

1 giving an upper bound for the maximum value

Obviously the larger the choice set, the larger the value of det(Copt,AfT)and hence the more efficient the design. For example, suppose there are 6 binary attributes. In Table 7.5 the values of det(Copt,nfT,m) and efficiencies are given for choice sets of size 2 to 10. The efficiency is calculated with respect to det(C,,t,h{T) with p = k k ( k - 1)/2. These results show that in this setting the size of the choice sets should be as large as the respondents can cope with.

+


Table 7.4 Efficiency of Different Values of Effects Only. m 2 3 4 5 6 7 8 9 10

II 12 13

14 15 16

71%.for

det(CnptAI 7.13073 x 6.33128 x 1.26912 x 1.00000 x 1.15137 x 1.19940 x 1.39920 x 1.32371 x 1.30573 x 1.34005 x 1.44397 x 1.36887 x 1.37192 x 1.39964 x 1.43271 x

Table 7.5 Efficiency of Different Values of Effects and Two-factor Interactions.

ni

Four Asymmetric Attributes for Main

m)

lo-*'

lo-'* lo-'* 10-Is 10-l' lo-'*

10-l' lo-'' 10-18

lo-'* lo-'* lo-''

Efficiency ('36)

7 1.59 91.25 98.58 96.00 97.52 97.96 99.65 99.04 98.89 99.17 100

99.4 1 99.43 99.65 99.9 I

for 6 Binary Attributes when Estimating Main

m

Max det(Cnfr)

Efficiency ('36)

2 3 4 5 6 7 8 9 10

9.25596 x 3.89166 x 4.61677 x 1 . 7 9 0 ~ 10-39 ,i.21935 x 10-39 7.62375 x 10-39 1.17549 x 1.6:3625 x 2.12395 x

50.00 66.67 75.00 80.00 83.33 85.7 1 87.50 88.89 90.00

~

241

242

VARIOUS TOPICS

7.2.3 Choice Experiments with Choice Sets of Various Sizes Although all the work on optimal choice experiments that has been done to date has assumed that all choice sets in a particular experiment are of the same size, there is in fact no mathematical requirement for that to be true. In this section we derive the information matrix for a forced choice experiment in which there are N , choice sets of size m,,, 1 5 s 5 T . We then show that the optimal choice set sizes are the ones that we have found in the previous sections. Extending the notation in the earlier chapters, we let n!:!z, be the number of times that options Y i l and Y i 2 appear together in the Ns choice sets of size 772,. Let N = N, be the total number of choice sets. Under the null hypothesis that all items are equally attractive, A contains the proportion of choice sets in which pairs of profiles appear together and hence

c,

-

A

1

=-

N

c r

N,5A,7.

s=l

Substituting, we get

Since the entries in the matrix for one choice set size have no effect on the entries in the matrix for another choice set size, we need to maximize the coefficients for each choice set size independently. If we are only estimating main effects, then from Theorem 7.2.1 we know that each m, needs to be a multiple of the least common multiple of B 1 , &, . . . ,e k to get the largest determinant. mEXAMPLE7.2.3. Suppose that there are k = 2 binary attributes and that we have choice sets of three different sizes: N1 = 2 choice sets, (00, 1 1 ) and (01, lo), each of size ml = 2 (design I ) ; N2 = 4 choice sets, (00,01, lo), (00, 01, 1 I ) , (00, 10, 1 1 ) and (01, 10, 1 I ) , each of size m2 = 3 (design 2); and N3 = 1 choice set, (00, 01, 10, 1 1 ) of size m3 = 4 (design 3). If we use just the choice sets of size 2 and 3 in the choice experiment (designs 1 & 2), then we = 0 and n&ill = nb:),lo = 1. (Recall that - noo,lo (1) = ol,ll - 7)(1) have that noo,ol (1) n(s) - n ( s ) .) For the sets of size 3, nig,)z,= 2 V i l # 22. The A matrix for these 6 2.1.72

92.91

choice sets is

A=

&+B

-_ -_ 1 - 1 _ 21_ 27 1 27 1 27

1 -27

&+f

1 _ 24_ _ 27 _I

1 27

1 _ 27_ 1 -_ - 27 1 24

A+; _-271

1 _ 24_1 - _ 27 1 - _27 - _271

A+;

-

-

If we wish to estimate main effects only, these 6 choice sets have an efficiency of 92.52% compared to the optimal experiments which are those that contain choice sets of size 2, or 4, or 2 and 4 together (see Table 7.6). Only using the 4 choice sets of size 3 (design 2) gives a design with an efficiency of 88.89% however: so including some choice sets of an optimal size improves the efficiency of the triples.

PARTIAL PROFILES

243

Table 7.6 Choice Sets of Up to Three Different Sizes for 6 Binary Attributes for Main Effects Only.

Design(s)

rn,

N

Max det(C.t,)

Values

Efficiency (%'r)

2 3

2 3 4

2 4 1

0.0625 0.0494 0.0625

I00 88.89

1 &2 1&3 2&3

2&3 2&4 3&4

6 3 5

0.0536 0.0625 0.05 I9

92.52 I00 9 1.04

1,2823

2.3&4

I

0.0548

93.58

1

100

7.2.4 Concluding Comments on Choice Set Size We have given results on the best choice set sizes to use to maximize the statistical efficiency of the choice experiment. Some authors, like DeShazo and Fenno (2002). have used larger choice set size as an indicator of a more complex task and so might argue that choice experiments in which several different choice set sizes are used have different values for respondent efficiency. Thus practitioners will need to decide how to trade-off gains i n statistical efficiency with potential losses i n respondent efficiency.

7.3 PARTIAL PROFILES If the number of attributes used to describe each option is of the order of 16 and if all the attributes are allowed to differ between the options in each choice set then there is evidence that respondents will not perform as consistently as they might; see, for instance, Swait and Adamowicz (1996), Severin (2000), and Maddala et al. (2002). Holling et al. (1998) provide a summary of a number of studies in which both the number of attributes to be used and the the number of choice sets to be answered was compared. More choice sets with more options slowed down respondents and decreased the predictive ability of the model. In Maddala et al. (2002), for example. the authors compared a choice experiment in which 6 attributes were allowed to differ between options in a choice set with a choice experiment in which only 4 of the 6 attributes were allowed to differ between the options in a choice set. They found no significant differences in consistency, perceived difficulty or fatigue but they did observe differences i n the stated preferences which these authors think may link with the importance of context to a stated preference experiment. Some authors have focused on finding optimal designs when there is a limit on the number of attributes that are different between options within a choice set. If only some of the attributes may differ between the options within a choice set, then the options are said to be described by a pnrtinl profile. One way to do this is to show all the attributes but have several attributes with the same level for all the options in a choice set. This was the approach taken by Severin (2000) using pairs. Another way to limit the number of attributes different is to choose subsets of the k attributes and find a choice experiment on the subset of attributes only; the remaining

244

VARIOUS TOPICS

attributes are not shown. Various such subsets are used in turn to get the complete choice experiment. This approach was introduced by Green (1974). If m = 2 , then allowing only f of the k attributes to be different, but showing all k attributes, means that the optimal design for estimating main effects is the design in which all choice sets with f attributes different are included. The corresponding det(Co,t.hf,pp) is given by putting a f = 1/(2k-' in the expression for det(Chf) in Theorem 4.1.1, This gives an optimal determinant of

(r))

Green (1974) used balanced incomplete block designs (BIBDs) (defined in Section 2.4) to determine which attributes would be allowed to vary in a set of choice sets. He used all the blocks of the BIBD in turn to get all of the choice sets in an experiment. WEXAMPLE 7.3.1.

Suppose that pizzas are described by the 16 attributes, each with two levels, as given in Table 7.7. Then a (16,20,5,4,1) BIBD (Table 7.8, from Mathon and Rosa (2006)) could be used to determine which four attributes were varied in the choice sets associated with each block. Each block of the BIBD would give rise to four choice sets, as illustrated in Table 7.9. Using all 20 x 4 = 80 choice sets gives a design which is 100% efficient if only four attributes are allowed to have different levels between the two options in a choice set and is 25% efficient if there are no restrictions on the number of attributes which may have different levels in the two options in a choice set. Kl Table 7.7 Sixteen Attributes Used to Describe Pizzas Attribute

Level 0

Level 1

Quality of ingredients Price Pizza temperature Manners of operator Delivery charge Delivery time Vegetarian available Pizza type Type of outlet Baking method Range Distance to outlet Type of crust Available sizes Opening hours

All fresh $13 Steaming hot Friendly Free 30 minutes Yes Traditional Chain Woodfire oven Large menu Same suburb Thin Single size Till 10 pm

Some tinned $17 warm Unfriendly $2 45 minutes No Gourmet Local Traditional oven Restricted menu Next suburb Thick Three sizes Till 1 am

Green's idea has been extended by using designs in which each pair of treatment combinations appears in at least one block by Grasshoff et al. (2004) rather than insisting on equality of pair replication as in the incomplete block design. Usually this relaxation means that the choice experiment is smaller but the relative efficiency of these two approaches has yet to be determined.

CHOICE EXPERIMENTS USING PRIOR POINT ESTIMATES

245

Table 7.8 (16.20,5,4,1)BIBD

0 1 2 3 O d e f 1 9 c f 2 8 a d

3 7 b e

0 1 2 3 4

4 4 4 4 8

5 7 c 9 b

6 a e d f

0 7 8 9 1 5 b d 2 5 7 f 3 5 8 c

5 9 a e

O 1 2 3 6

a 6 6 6 7

b 8 9 a c

c e b f d

Table 7.9 The Four Choice Sets from the First Block of the (16, 20, 5.4, I)BIHI) Option 1

Fresh $13 Hot

Option 2

Friendly

Tinned $17 Warm Unfriendly

Fresh $13 Warm Unfriendly

Tinned $17 Hot

Fresh $17 Hot

Tinned $13 Warm Friendly

Unfriendly

Fresh $17 Warin Friendly

Tinned $13 Hot

Friendly

Unfriendly

7.4 CHOICE EXPERIMENTS USING PRIOR POINT ESTIMATES So far i n this book we have constructed optimal designs under the null hypothesis of no difference between the options. However the expression for the i l matrix given in Equations (3.2) and (3.3) includes the 7ri values and so in fact we could find the optimal designs for any prior values of the ir, that we thought were appropriate. We discuss this idea in this section. Additional information about the relative magnitudes of the 7r7 values may be available from earlier studies or from taking a sequential approach to experimentation. Wc give two examples below. In the first there is a wide range of values of 7r where the optimal designs for the null hypothesis perform well. In the second case the optimal designs for the null hypothesis are at best about 63% efficient. In the first example we consider two binary attributes and let the class of competing designs be any set of pairs. This means that we have extended the class of competing designs from what we have considered i n the remainder of the book. This is possible because the computations are not overwhelming in this small case. We have assumed that the two-factor interaction effect is 0 so that we can work with the simplest form of the C matrix. EXAMPLE 7.4.1.

Suppose that there are k = 2 binary attributes and that the choice sets are of size m = 2. There are 4 treatment combinations and 6 pairs of distinct treatment combinations. Each such pair could be a choice set in the experiment. Suppose that all 26 - 1 = 63 non-empty sets of distinct pairs form the class of competing designs. Suppose also that i r 1 = I//?, 7r2 = ir:

The Construction of Optimal Stated Choice Experiments: Theory and Methods (Wiley Series in Probability and Statistics)

The Construction of Optimal Stated Choice Experiments

Alternative Methods of Regression (Wiley Series in Probability and Statistics)

Methods of Multivariate Analysis (Wiley Series in Probability and Statistics)

Probability and Measure (Wiley Series in Probability and Statistics)

Modern Regression Methods (Wiley Series in Probability and Statistics)

Introduction to Combinatorial Theory (Wiley Series in Probability and Statistics)

Stated choice methods: analysis and applications

Statistics: A Biomedical Introduction (Wiley Series in Probability and Statistics)

Spatial Statistics (Wiley Series in Probability and Statistics)

Statistics for Spatial Data (Wiley Series in Probability and Statistics)

Statistical Factor Analysis and Related Methods: Theory and Applications (Wiley Series in Probability and Statistics)

Applied Econometric Time Series (Wiley Series in Probability and Statistics)

Multiple Time Series (Wiley Series in Probability and Mathematical Statistics)

Models for Probability and Statistical Inference: Theory and Applications (Wiley Series in Probability and Statistics)

Statistical Rules of Thumb (Wiley Series in Probability and Statistics)

Applied Logistic Regression (Wiley Series in Probability and Mathematical Statistics. Applied Probability and Statistics Section)

An Introduction to Probability and Statistics (Wiley Series in Probability and Statistics)

Alternative methods of regression (Wiley Series in Probablity and Statistics)

Experiments: Planning, Analysis, and Optimization (Wiley Series in Probability and Statistics)

The Theory and Practice of Econometrics (Wiley Series in Probability and Statistics)

Operator-Limit Distributions in Probability Theory (Wiley Series in Probability and Statistics)

Nonlinear Regression (Wiley Series in Probability and Statistics)

The Construction of Optimal Stated Choice Experiments: Theory and Methods (Wiley Series in Probability and Statistics)